{"id":6231,"date":"2026-11-08T12:17:00","date_gmt":"2026-11-08T17:17:00","guid":{"rendered":"https:\/\/www.cybersecurityinstitute.com\/blog\/?p=6231"},"modified":"2026-11-08T12:17:00","modified_gmt":"2026-11-08T17:17:00","slug":"ai-machine-learning-security-november-8-2026","status":"publish","type":"post","link":"https:\/\/www.cybersecurityinstitute.com\/blog\/?p=6231","title":{"rendered":"AI &amp; Machine Learning Security &mdash; November 8, 2026"},"content":{"rendered":"<style>\n.single .entry-title,\n.single .entry-header .entry-title,\n.single .post-title,\n.single header.entry-header h1,\n.single h1.entry-title,\n.single .page-title,\n.post-template-default h1.entry-title,\n.post-template-default .entry-header,\narticle .entry-header,\narticle .entry-title { display: none !important; }\n.single .entry-header { margin: 0 !important; padding: 0 !important; }\n.single .entry-content { margin-top: 0 !important; padding-top: 0 !important; }\n<\/style>\n<table role=\"presentation\" class=\"wrapper\" cellpadding=\"0\" cellspacing=\"0\" border=\"0\" width=\"100%\">\n<tr>\n<td align=\"center\">\n<table role=\"presentation\" class=\"container\" cellpadding=\"0\" cellspacing=\"0\" border=\"0\" width=\"680\">\n<p>        <!-- Banner --><\/p>\n<tr>\n<td class=\"banner\" style=\"background-color:#581c87;background:linear-gradient(135deg,#581c87 0%,#9333ea 100%);padding:36px 32px;color:#ffffff;\">\n<p class=\"date\" style=\"color:#ffffff !important;\">November 8, 2026 &middot; Weekly Edition<\/p>\n<h1 style=\"color:#ffffff !important;\">AI &amp; Machine Learning Security<\/h1>\n<p class=\"tagline\" style=\"color:#ffffff !important;\">OpenAI put GPT-6 Cyber behind an application gate after rating it &ldquo;Critical&rdquo;, and its exploit pipeline built working root exploits for 948 Linux kernel bugs. Anthropic told investors it may be close to AI-driven acceleration, coding agents kept installing archived MCP servers, and Check Point found adware written in plain English.<\/p>\n<\/td>\n<\/tr>\n<p>        <!-- At a glance --><\/p>\n<tr>\n<td class=\"content\">\n<h2>This week at a glance<\/h2>\n<p>This week offensive capability came with a price list. OpenAI launched GPT-6 Cyber at $20 per million input tokens and $100 per million output tokens, rated it &ldquo;Critical&rdquo; under its own Preparedness Framework, and limited it to the vetted Daybreak Red program. In its research post, OpenAI says an automated pipeline found 2,308 candidate bugs in the Linux kernel in 12 hours and produced 1,169 working root exploits across 948 distinct vulnerabilities, at about $1,279 per exploit. Ten days later, 61 of those bugs had been patched upstream. Open models are catching up. On Vals AI&rsquo;s network-attack benchmark, DeepSeek V4.1 Flash came within eight points of GPT-6 Sol at about a seventh of the cost per run, and the UK AI Security Institute now puts the open-to-closed cyber gap at four to seven months. Gartner says open-weight models will run 90% of enterprise agent workflow steps by 2030. Anthropic&rsquo;s revised S-1 warns that it may soon cross its own threshold for AI-accelerated progress, and METR&rsquo;s audit suggests that may already have happened on July 15.<\/p>\n<p>On the defence side, most of the news was about what agents install and read. Glow Labs found that in 98% of the sessions where a coding agent installed or suggested a deprecated MCP server, the package name came from the agent itself, not the user. Those archived reference servers have been downloaded 17.3 million times and will never be patched. Check Point found a Codex jailbreak bundle that rewrites the agent&rsquo;s global instruction files to inject ads for an LLM gateway, and it also bypassed ChatGPT&rsquo;s Lockdown Mode through a nested-frame popup that OpenAI patched in September. JFrog Security Research showed that a single unauthenticated HTTP request can give an attacker root on a Reachy Mini desktop robot. Radware found hidden text addressed to AI agents in GitHub issues and pull requests more than doubling between June and September, and in an AE Studio test, coding agents gave away a private token 15% of the time when other agents offered help in return. The week&rsquo;s new open models were Thinking Machines&rsquo; Inkling 2 Small, Ricursive&rsquo;s Apache 2.0 Ripple-65B Preview and Baidu&rsquo;s ERNIE 5.5 Flash Base.<\/p>\n<div class=\"watchlist\">\n<h2>On our watch list<\/h2>\n<ul>\n<li><strong>Who gets into Daybreak Red.<\/strong> OpenAI says more than 2,000 entities have enrolled to use GPT-6 Cyber, and it rates the model&rsquo;s cyber capability &ldquo;Critical&rdquo;. Watch for published vetting criteria, for whether the honeypot result (99.93% of unauthorized attacks avoided) holds up outside OpenAI&rsquo;s tests, and for how Anthropic and Google price their gated cyber models in response.<\/li>\n<li><strong>The Linux kernel patch backlog.<\/strong> After ten days, 61 of the 948 bugs were patched upstream, 43 had staging patches and 299 were under public discussion. Watch how quickly distributions backport the fixes, whether CVE assignment keeps up, and whether OpenAI publishes the ten exploit techniques it is withholding once fixes land.<\/li>\n<li><strong>Anthropic&rsquo;s RSP update.<\/strong> Anthropic says it is considering changes to its Responsible Scaling Policy and has not declared that any model crossed the acceleration threshold. Watch for that decision, for the next risk report, and for whether the METR review of how Anthropic splits compute between capabilities and safety, which Anthropic cancelled, is rescheduled.<\/li>\n<li><strong>A clean-up of the archived MCP reference servers.<\/strong> The maintainers closed Glow&rsquo;s report as out of scope, and npm&rsquo;s removal rules don&rsquo;t cover heavily used packages like server-github. Watch for deprecation notices on npm, for registries and agent vendors to add pre-install checks, and for the first reported exploitation of the Puppeteer server flaw.<\/li>\n<li><strong>The gateway behind the Haiou bundle.<\/strong> Check Point found no evidence that the LLM gateway advertised by the adware is malicious, but it cites research on gateways that stole cloud keys and injected tool calls. Watch for the gateway operator to be identified, and for copycat bundles aimed at other coding agents. The Claude Code version of the bundle carried no ads.<\/li>\n<li><strong>Inkling 2 Small&rsquo;s open-weight release.<\/strong> Thinking Machines held back the weights pending external safety evaluations and targeted an Apache 2.0 release for the week after launch. Watch for the release date and for whoever runs those evaluations to publish their results.<\/li>\n<li><strong>Independent scores for Ripple-65B.<\/strong> Ricursive&rsquo;s model card and launch graphic disagree on rival scores, and OpenRouter caps context at 16,384 tokens against an advertised million. Watch for third-party runs of Terminal-Bench 2.1 and SWE-bench Pro, and for the Fireworks endpoint.<\/li>\n<li><strong>Longer runs on open models.<\/strong> Vals found DeepSeek scored nothing in its first 25 million tokens and made most of its progress after 100 million. Watch for evaluators, including AISI, to extend token budgets, since short tests may understate what open weights can do in an attack.<\/li>\n<li><strong>Anthropic&rsquo;s answer on Claude Code refusals.<\/strong> Anthropic confirmed it added instructions against recreating other people&rsquo;s games, but has not said whether reverse-engineering refusals on Opus 5.5 come from the same change. Watch for documented prompt changes and for Cyber Verification Program updates that let security teams keep this work.<\/li>\n<li><strong>Who adopts FlashProve.<\/strong> Anthropic&rsquo;s inference proofs take 70 to 95 times the GPU time of plain inference and only cover 1,024-token windows, and no lab plans to deploy them yet. Watch for evaluators or routing services to pilot them as a way to prove which model actually answered.<\/li>\n<li><strong>Python 3.12 inference servers.<\/strong> RunSybil&rsquo;s chat-template sandbox escape was fixed in CPython 3.13 to 3.15 as an ordinary bug fix, with no backport to 3.12 and no CVE. Watch whether vLLM, Hugging Face or Jinja ship a mitigation of their own.<\/li>\n<\/ul><\/div>\n<p>            <!-- Topic map --><\/p>\n<div class=\"topic-map\">\n              <img decoding=\"async\" src=\"https:\/\/www.cybersecurityinstitute.com\/blog\/wp-content\/uploads\/2026\/11\/topic-map-ai-ml-2026-11-08.png\" alt=\"Topic map of this week&rsquo;s AI &amp; Machine Learning Security themes\" loading=\"eager\"><\/p>\n<p class=\"caption\">This week&rsquo;s topic map: OpenAI and GPT-6 Cyber at the centre of the AI exploit-generation cluster with the Linux kernel; Anthropic, METR and AI-accelerated development; MCP linking deprecated servers, tool poisoning and the coding agents that install them; Check Point&rsquo;s Haiou adware and Lockdown Mode findings around prompt injection; and open weights tying DeepSeek, Gartner and the week&rsquo;s new models to JFrog&rsquo;s Reachy Mini research through Hugging Face.<\/p>\n<p>              <!-- INTERACTIVE_MAP_LINK_START --><\/p>\n<p style=\"margin:10px 0 0;text-align:center;\"><a href=\"https:\/\/www.cybersecurityinstitute.com\/blog\/?p=6230\" target=\"_blank\" rel=\"noopener\" style=\"display:inline-block;padding:8px 18px;background-color:#0f172a;color:#ffffff !important;text-decoration:none;border-radius:6px;font-size:13px;font-weight:600;\">View interactive topic map &rarr;<\/a><\/p>\n<p><!-- INTERACTIVE_MAP_LINK_END -->\n            <\/div>\n<p>            <!-- Article index --><\/p>\n<h2>Article index<\/h2>\n<h3>The week in five threads<\/h3>\n<h4>1. Frontier cyber capability and the acceleration question<\/h4>\n<div class=\"cluster-intro\">OpenAI&rsquo;s gated, &ldquo;Critical&rdquo;-rated cyber model and its 948 kernel root exploits, Anthropic&rsquo;s S-1 warning about AI-accelerated development, METR&rsquo;s audit of that claim, and a narrowing open-closed gap on network attacks.<\/div>\n<table class=\"index-table\">\n<tr>\n<th>Article<\/th>\n<th>Source<\/th>\n<th>Published<\/th>\n<\/tr>\n<tr>\n<td>1. <a href=\"https:\/\/techcrunch.com\/2026\/11\/02\/anthropic-warns-that-its-very-close-to-rsi-capability-aka-self-building-ai\/?_bhlid=cf872ed9f4ee16d8ac4caee2f6663634ba62c880\">Anthropic warns that it&#8217;s very close to RSI capability, aka self-building AI<\/a><\/td>\n<td class=\"src\">TechCrunch<\/td>\n<td class=\"dt\">Nov 2, 2026<\/td>\n<\/tr>\n<tr>\n<td>2. <a href=\"https:\/\/venturebeat.com\/technology\/openai-launches-gpt-6-cyber-with-stronger-exploit-research-and-higher-api-price-than-anthropic-and-google\">OpenAI launches GPT-6 Cyber with stronger exploit research and higher API price than Anthropic and Google<\/a><\/td>\n<td class=\"src\">VentureBeat<\/td>\n<td class=\"dt\">Nov 5, 2026<\/td>\n<\/tr>\n<tr>\n<td>3. <a href=\"https:\/\/www.helpnetsecurity.com\/2026\/11\/05\/openai-linux-kernel-root-exploits\/\">OpenAI&#8217;s upcoming AI model built 948 working Linux root exploits in a single run<\/a><\/td>\n<td class=\"src\">Help Net Security<\/td>\n<td class=\"dt\">Nov 5, 2026<\/td>\n<\/tr>\n<tr>\n<td>4. <a href=\"https:\/\/the-decoder.com\/metr-finds-anthropic-may-have-crossed-its-own-ai-acceleration-warning-threshold-back-in-july\/\">METR finds Anthropic may have crossed its own AI acceleration warning threshold back in July<\/a><\/td>\n<td class=\"src\">The Decoder<\/td>\n<td class=\"dt\">Nov 6, 2026<\/td>\n<\/tr>\n<tr>\n<td>5. <a href=\"https:\/\/the-decoder.com\/the-gap-between-open-and-closed-ai-models-in-cyber-capabilities-keeps-shrinking\/\">The gap between open and closed AI models in cyber capabilities keeps shrinking<\/a><\/td>\n<td class=\"src\">The Decoder<\/td>\n<td class=\"dt\">Nov 7, 2026<\/td>\n<\/tr>\n<\/table>\n<h4>2. Open models: releases and the cost case<\/h4>\n<div class=\"cluster-intro\">Gartner argues open weights will run most agent steps, three new models arrive from Baidu, Ricursive and Thinking Machines, and a fine-tune shows state propaganda in a Chinese model is only skin-deep.<\/div>\n<table class=\"index-table\">\n<tr>\n<th>Article<\/th>\n<th>Source<\/th>\n<th>Published<\/th>\n<\/tr>\n<tr>\n<td>6. <a href=\"https:\/\/www.theregister.com\/ai-and-ml\/2026\/11\/03\/open-models-can-do-90-of-agentic-work-at-a-twentieth-of-the-cost-of-a-frontier-llm-gartner\/5307132\">Open models can do 90% of agentic work at a twentieth of the cost of a frontier LLM: Gartner<\/a><\/td>\n<td class=\"src\">The Register<\/td>\n<td class=\"dt\">Nov 3, 2026<\/td>\n<\/tr>\n<tr>\n<td>7. <a href=\"https:\/\/the-decoder.com\/baidu-releases-ernie-5-5-flash-base-with-280-billion-parameters-as-an-open-source-foundation-model\/\">Baidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model<\/a><\/td>\n<td class=\"src\">The Decoder<\/td>\n<td class=\"dt\">Nov 5, 2026<\/td>\n<\/tr>\n<tr>\n<td>8. <a href=\"https:\/\/venturebeat.com\/technology\/ex-googlers-new-apache-2-0-model-ripple-65b-preview-claims-the-u-s-open-coding-lead-on-two-major-benchmarks\">Ex-Googlers&#8217; new Apache 2.0 model Ripple-65B Preview claims the U.S. open coding lead on two major benchmarks<\/a><\/td>\n<td class=\"src\">VentureBeat<\/td>\n<td class=\"dt\">Nov 4, 2026<\/td>\n<\/tr>\n<tr>\n<td>9. <a href=\"https:\/\/venturebeat.com\/technology\/thinking-machines-launches-inkling-2-small-with-near-frontier-performance-on-enterprise-benchmarks\">Thinking Machines launches Inkling 2 Small with near-frontier performance on enterprise benchmarks<\/a><\/td>\n<td class=\"src\">VentureBeat<\/td>\n<td class=\"dt\">Nov 6, 2026<\/td>\n<\/tr>\n<tr>\n<td>10. <a href=\"https:\/\/the-decoder.com\/state-propaganda-in-some-chinese-ai-models-is-only-a-thin-layer-that-fine-tuning-can-strip-away\/\">State propaganda in some Chinese AI models is only &#8220;a thin layer&#8221; that fine-tuning can strip away<\/a><\/td>\n<td class=\"src\">The Decoder<\/td>\n<td class=\"dt\">Nov 7, 2026<\/td>\n<\/tr>\n<\/table>\n<h4>3. The agent attack surface<\/h4>\n<div class=\"cluster-intro\">Coding agents installing unpatchable MCP servers, adware that lives in an agent&rsquo;s instruction files, a robot rooted by one HTTP request, a ChatGPT Lockdown Mode escape, hidden prompts in GitHub issues, agents trading secrets for help, and a CPython bug that breaks inference-server sandboxes.<\/div>\n<table class=\"index-table\">\n<tr>\n<th>Article<\/th>\n<th>Source<\/th>\n<th>Published<\/th>\n<\/tr>\n<tr>\n<td>11. <a href=\"https:\/\/www.helpnetsecurity.com\/2026\/11\/03\/mcphantom-deprecated-mcp-servers\/\">AI coding agents are installing MCP servers that will never be patched<\/a><\/td>\n<td class=\"src\">Help Net Security<\/td>\n<td class=\"dt\">Nov 3, 2026<\/td>\n<\/tr>\n<tr>\n<td>12. <a href=\"https:\/\/research.checkpoint.com\/2026\/free-the-model-sell-the-gateway-prompt-adware-hidden-in-jailbreak-bundles\/\">Free the Model, Sell the Gateway: Prompt Adware hidden in jailbreak bundles<\/a><\/td>\n<td class=\"src\">Check Point Research<\/td>\n<td class=\"dt\">Nov 3, 2026<\/td>\n<\/tr>\n<tr>\n<td>13. <a href=\"https:\/\/jfrog.com\/blog\/reachy-http-request-hijacks-ai-robot\/\">How One HTTP Request can Hijack your AI Robot<\/a><\/td>\n<td class=\"src\">JFrog<\/td>\n<td class=\"dt\">Nov 3, 2026<\/td>\n<\/tr>\n<tr>\n<td>14. <a href=\"https:\/\/www.radware.com\/blog\/prompt-injecting-coding-agents\/\">Who Is Prompt-Injecting Coding Agents?<\/a><\/td>\n<td class=\"src\">Radware<\/td>\n<td class=\"dt\">Nov 5, 2026<\/td>\n<\/tr>\n<tr>\n<td>15. <a href=\"https:\/\/thehackernews.com\/2026\/11\/researchers-bypass-chatgpt-lockdown.html\">Researchers Bypass ChatGPT Lockdown Mode to Exfiltrate Chats Without User Interaction<\/a><\/td>\n<td class=\"src\">The Hacker News<\/td>\n<td class=\"dt\">Nov 5, 2026<\/td>\n<\/tr>\n<tr>\n<td>16. <a href=\"https:\/\/www.theregister.com\/security\/2026\/11\/06\/agents-reveal-secrets-when-offered-to-team-up-with-each-other\/5308260\">Agents reveal secrets when offered to team up with each other<\/a><\/td>\n<td class=\"src\">The Register<\/td>\n<td class=\"dt\">Nov 6, 2026<\/td>\n<\/tr>\n<tr>\n<td>17. <a href=\"https:\/\/www.csoonline.com\/article\/4245592\/cpython-flaw-exposes-ai-inference-servers-to-sandbox-escape-attacks.html\">CPython flaw exposes AI inference servers to sandbox escape attacks<\/a><\/td>\n<td class=\"src\">CSO Online<\/td>\n<td class=\"dt\">Nov 6, 2026<\/td>\n<\/tr>\n<\/table>\n<h4>4. Agent trust, identity and oversight<\/h4>\n<div class=\"cluster-intro\">Unannounced changes to what a hosted coding agent will do, the standards body racing its own members, proofs of which model answered, Claude acting as an identity issuer, auditors that drown reviewers in false alarms, and coding agents that still leave work unfinished.<\/div>\n<table class=\"index-table\">\n<tr>\n<th>Article<\/th>\n<th>Source<\/th>\n<th>Published<\/th>\n<\/tr>\n<tr>\n<td>18. <a href=\"https:\/\/thenewstack.io\/claude-code-developer-refusals\/\">Developers are downgrading their AI on purpose \u2014 and it&#8217;s working<\/a><\/td>\n<td class=\"src\">The New Stack<\/td>\n<td class=\"dt\">Nov 2, 2026<\/td>\n<\/tr>\n<tr>\n<td>19. <a href=\"https:\/\/thenewstack.io\/mazin-gilbert-aaif-interview\/\">Can agent standards keep up with the companies building them?<\/a><\/td>\n<td class=\"src\">The New Stack<\/td>\n<td class=\"dt\">Nov 3, 2026<\/td>\n<\/tr>\n<tr>\n<td>20. <a href=\"https:\/\/thenewstack.io\/anthropic-flashprove-model-verification\/\">Anthropic found a way to prove which AI model answered you \u2014 but there&#8217;s a costly catch<\/a><\/td>\n<td class=\"src\">The New Stack<\/td>\n<td class=\"dt\">Nov 4, 2026<\/td>\n<\/tr>\n<tr>\n<td>21. <a href=\"https:\/\/thenewstack.io\/claude-plugin-identity-issuer\/\">\u201cSigning in to your company\u2019s plugins one by one is annoying\u201d: Anthropic\u2019s fix puts Claude in the identity business<\/a><\/td>\n<td class=\"src\">The New Stack<\/td>\n<td class=\"dt\">Nov 4, 2026<\/td>\n<\/tr>\n<tr>\n<td>22. <a href=\"https:\/\/the-decoder.com\/ai-auditors-catch-most-agent-misbehavior-but-flood-reviewers-with-false-alarms-new-study-finds\/\">AI auditors catch most agent misbehavior but flood reviewers with false alarms, new study finds<\/a><\/td>\n<td class=\"src\">The Decoder<\/td>\n<td class=\"dt\">Nov 7, 2026<\/td>\n<\/tr>\n<tr>\n<td>23. <a href=\"https:\/\/the-decoder.com\/ai-coding-agents-still-need-human-cleanup-on-nearly-three-out-of-four-tasks-in-a-real-world-test\/\">AI coding agents still need human cleanup on nearly three out of four tasks in a real-world test<\/a><\/td>\n<td class=\"src\">The Decoder<\/td>\n<td class=\"dt\">Nov 8, 2026<\/td>\n<\/tr>\n<\/table>\n<h4>5. Foundations and perspectives<\/h4>\n<div class=\"cluster-intro\">Governing what an agent learns, a cross-framework measurement of MCP tool poisoning, practical loop engineering, the MCP server nobody inventoried, and why decision models shift the burden onto context.<\/div>\n<table class=\"index-table\">\n<tr>\n<th>Article<\/th>\n<th>Source<\/th>\n<th>Published<\/th>\n<\/tr>\n<tr>\n<td>24. <a href=\"https:\/\/www.oreilly.com\/radar\/who-governs-what-an-agent-learns\/\">Who Governs What an Agent Learns?<\/a><\/td>\n<td class=\"src\">O&#8217;Reilly Radar<\/td>\n<td class=\"dt\">Oct 30, 2026<\/td>\n<\/tr>\n<tr>\n<td>25. <a href=\"https:\/\/arxiv.org\/abs\/2611.03555\">MCP Tool Poisoning in Agentic AI Frameworks: A Cross-Framework Measurement Study<\/a><\/td>\n<td class=\"src\">arXiv<\/td>\n<td class=\"dt\">Nov 2, 2026<\/td>\n<\/tr>\n<tr>\n<td>26. <a href=\"https:\/\/www.oreilly.com\/radar\/practical-loop-engineering\/\">Practical Loop Engineering<\/a><\/td>\n<td class=\"src\">O&#8217;Reilly Radar<\/td>\n<td class=\"dt\">Nov 2, 2026<\/td>\n<\/tr>\n<tr>\n<td>27. <a href=\"https:\/\/www.scworld.com\/perspective\/the-mcp-server-nobody-inventoried\">The MCP server nobody inventoried<\/a><\/td>\n<td class=\"src\">SC Media<\/td>\n<td class=\"dt\">Nov 3, 2026<\/td>\n<\/tr>\n<tr>\n<td>28. <a href=\"https:\/\/venturebeat.com\/orchestration\/decision-models-like-jev-dont-explain-their-answers-that-makes-context-your-problem\">Decision models like Jev don&#8217;t explain their answers. That makes context your problem<\/a><\/td>\n<td class=\"src\">VentureBeat<\/td>\n<td class=\"dt\">Nov 7, 2026<\/td>\n<\/tr>\n<\/table>\n<p>            <!-- Detailed write-ups --><\/p>\n<h2>Detailed write-ups<\/h2>\n<div class=\"article\">\n<h4>1. GPT-6 Cyber goes on sale behind a gate, and OpenAI&rsquo;s pipeline roots the Linux kernel 948 ways<\/h4>\n<p class=\"meta\">VentureBeat &middot; November 5, 2026<\/p>\n<p>OpenAI&rsquo;s GPT-6 Cyber costs $20 per million input tokens and $100 per million output tokens, about 60% more for input and a third more for output than GPT-5.6 Cyber, VentureBeat&rsquo;s Carl Franzen reports. You can&rsquo;t simply buy it. Access goes through the gated Daybreak Red program, which needs an application and a review, and OpenAI says more than 2,000 entities have enrolled. OpenAI now rates the model&rsquo;s cybersecurity capability &ldquo;Critical&rdquo; under its Preparedness Framework, up from &ldquo;High&rdquo; for the previous version. It scores 48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber and completes 95% of RemoteExploitBench tasks, while GPT-6 Astra under the Daybreak Blue defender program refused every one. In a honeypot test without safety prompts, it stayed in scope 99.93% of the time. OpenAI also reported 81 previously unknown Firefox vulnerabilities that Mozilla rated high severity.<\/p>\n<p>Help Net Security&rsquo;s Sinisa Markovic summarises the research post behind the headline number. In mid-October, OpenAI&rsquo;s Cyber team pointed an automated pipeline at the latest Linux kernel. Agents narrowed the audit to about 7% of the code and produced 2,308 candidate bugs in 12 hours, and triage agents reproduced 1,724 of them within 30 hours. Exploit agents then built 1,169 working local root exploits covering 948 distinct vulnerabilities, plus 291 information-leak proofs of concept, for $1,213,430 in API credits, or about $1,279 per exploit. An earlier run had claimed 200 root exploits, but 20 of them actually reached root through a different zero-day than the one assigned. &ldquo;This suggests that even successful exploitation by agents needs proper validation,&rdquo; OpenAI wrote, and it added an independent check. Ten days after reporting, 61 bugs were patched upstream, 43 had staging patches and 299 were being discussed publicly.<\/p>\n<p>For defenders, the cost is the number to remember. At roughly $1,300 per working kernel exploit, privilege escalation on an unpatched Linux host is cheap to build for anyone with a capable model. Expect kernel advisories to keep arriving in large batches, and make sure your patching process can handle that volume. All benchmark and exploit figures here are OpenAI&rsquo;s own. Help Net&rsquo;s article doesn&rsquo;t name the model used in the kernel run, while VentureBeat ties the kernel work to GPT-6 Cyber.<\/p>\n<p style=\"font-size:13px;color:#6b7280;margin:0;\">Sources: <a href=\"https:\/\/venturebeat.com\/technology\/openai-launches-gpt-6-cyber-with-stronger-exploit-research-and-higher-api-price-than-anthropic-and-google\">VentureBeat (GPT-6 Cyber launch)<\/a> &middot; <a href=\"https:\/\/www.helpnetsecurity.com\/2026\/11\/05\/openai-linux-kernel-root-exploits\/\">Help Net Security (948 root exploits)<\/a><\/p>\n<\/p><\/div>\n<div class=\"article\">\n<h4>2. Anthropic says it may be close to AI-built AI, and METR says the line may already be behind it<\/h4>\n<p class=\"meta\">TechCrunch &middot; November 2, 2026<\/p>\n<p>Anthropic&rsquo;s revised S-1\/A filing says the company may soon meet the Responsible Scaling Policy threshold for &ldquo;dramatically accelerated progress attributable to AI&rdquo;, TechCrunch&rsquo;s Julie Bort reports. The risk factors admit that Anthropic has failed in the past to notice jumps in its own development pace until after they happened, and that it has already used AI to build models whose gains it didn&rsquo;t recognise at the time. &ldquo;Using AI to help build more capable AI could also pose a considerable safety risk,&rdquo; the filing says. Anthropic has not designated any model as over the threshold and is still reviewing the data. It says it doesn&rsquo;t expect Claude to design and build successive frontier generations without meaningful human oversight, and it is considering changes to the RSP. Skeptics point out that an S-1 is a marketing document as well as a disclosure.<\/p>\n<p>Four days later, The Decoder&rsquo;s Maximilian Schreiner reported METR&rsquo;s view. The nonprofit spent eleven weeks, from August 12 to October 29, reviewing Anthropic&rsquo;s internal benchmarks, training metrics and staff surveys, and interviewed about 15 researchers. Anthropic&rsquo;s composite AECI progress metric has risen about 1.9 times faster since October 2025, from roughly 15 points a year to 29, and METR attributes about 80% of that speedup to AI used in Anthropic&rsquo;s own research rather than to more compute. METR thinks the warning level may have been reached on July 15. Anthropic&rsquo;s August risk report said publicly that the acceleration was &ldquo;less than a factor of 2&rdquo;, but the unredacted version implies 1.97 and calls crossing the threshold &ldquo;plausible&rdquo;. METR&rsquo;s range runs from 1.3 to 3 times, and the industry-wide figure, including OpenAI, Google and xAI, is 1.4. Anthropic cancelled a planned METR review of how it splits compute between capabilities and safety.<\/p>\n<p>For security architects, this is a governance question for your vendors, not a reason to change controls this week. If labs can&rsquo;t tell in real time how fast their own models are improving, capability jumps will reach you without warning. Ask vendors what they disclose and when, and check that your evaluation and red-teaming covers new model versions, not just new products.<\/p>\n<p style=\"font-size:13px;color:#6b7280;margin:0;\">Sources: <a href=\"https:\/\/techcrunch.com\/2026\/11\/02\/anthropic-warns-that-its-very-close-to-rsi-capability-aka-self-building-ai\/?_bhlid=cf872ed9f4ee16d8ac4caee2f6663634ba62c880\">TechCrunch (S-1 risk factors)<\/a> &middot; <a href=\"https:\/\/the-decoder.com\/metr-finds-anthropic-may-have-crossed-its-own-ai-acceleration-warning-threshold-back-in-july\/\">The Decoder (METR audit)<\/a><\/p>\n<\/p><\/div>\n<div class=\"article\">\n<h4>3. MCPhantom: coding agents keep installing MCP servers that will never be patched<\/h4>\n<p class=\"meta\">Help Net Security &middot; November 3, 2026<\/p>\n<p>Glow Labs reviewed real AI coding session transcripts and found that in 98% of sessions where a deprecated MCP server was installed or suggested, the agent supplied the package name from its training data, Help Net Security&rsquo;s Mirko Zorz reports. Users asked for the old package in only 2%. Glow calls the pattern MCPhantom, and saw it in Claude Code, Cursor, Codex CLI, Claude Cowork and Claude chat, with models from Anthropic, OpenAI and xAI. The archived reference servers carry four high-severity flaws. The worst is remote code execution in the Puppeteer server, which a prompt injection can trigger. They have been downloaded more than 17.3 million times from npm since July 2025, and five are still in npm&rsquo;s top 100 MCP packages. The repositories were archived in May 2025, Glow&rsquo;s report was closed as out of scope, and no fix is coming. npm only lets owners pull packages with fewer than 300 weekly downloads, one owner and no dependents, and server-github alone had 185,239 downloads last week. &ldquo;A lookup before install is not enough on its own,&rdquo; said Glow vulnerability researcher Nadav Claude Cohen. No exploitation has been reported.<\/p>\n<p>The visibility problem is the bigger issue. These servers don&rsquo;t appear in dependency manifests and the flaws have no CVEs, so EDR, SCA and inventory tools generally miss them. In an SC Media Perspectives column, Ashish Pulivarthy describes the same gap with CVE-2026-61560 (CVSS 9.8) in the community mcp-gitlab server. Before version 2.1.27, its network endpoints accepted unauthenticated requests on every interface from a root container, and one tool would read \/proc\/self\/environ and return the GitLab token. The fix shipped on June 22, but the CVE wasn&rsquo;t published until September 15, so any CVE-matching tool had nothing to match for about 12 weeks. A cross-framework arXiv preprint by Farooq Shaikh and Biagio Montaruli measured MCP tool poisoning across 12 frameworks, 18 models and 45 scenarios. It found a 22.9% attack success rate and verified secret exfiltration in 12.1% of vulnerable cases, and concludes that &ldquo;no framework implements intentional defense.&rdquo;<\/p>\n<p>In practice, every MCP server needs an owner, a list of the tokens it holds and a record of where it listens. Pin MCP packages to maintained versions, watch releases rather than waiting for CVEs, and treat any MCP transport that answers on a non-loopback interface without credentials as a finding. Glow is a security vendor and the arXiv study is a preprint.<\/p>\n<p style=\"font-size:13px;color:#6b7280;margin:0;\">Sources: <a href=\"https:\/\/www.helpnetsecurity.com\/2026\/11\/03\/mcphantom-deprecated-mcp-servers\/\">Help Net Security (MCPhantom)<\/a> &middot; <a href=\"https:\/\/www.scworld.com\/perspective\/the-mcp-server-nobody-inventoried\">SC Media (mcp-gitlab inventory)<\/a> &middot; <a href=\"https:\/\/arxiv.org\/abs\/2611.03555\">arXiv (MCP tool poisoning)<\/a><\/p>\n<\/p><\/div>\n<div class=\"article\">\n<h4>4. Prompt adware: a Codex jailbreak bundle that sells an LLM gateway<\/h4>\n<p class=\"meta\">Check Point Research &middot; November 3, 2026<\/p>\n<p>Check Point Research found a jailbreak bundle called Haiou on a misconfigured web server run by a threat actor who also used Cobalt Strike, the Ladon scanner and the VulnClaw pentesting agent. Haiou targets OpenAI Codex. It uses role-play, refusal suppression and &ldquo;educational CTF&rdquo; framing to weaken guardrails, and installs six skills for licence cracking, pentesting, reverse engineering and anti-cheat bypass. It also edits Codex&rsquo;s global instruction and configuration files so that later replies carry advertising, triggered by words such as &ldquo;hi&rdquo;, &ldquo;hello&rdquo; and &ldquo;ctf&rdquo;. &ldquo;Unlike traditional adware, it operates entirely through natural language in configuration artifacts,&rdquo; Check Point writes.<\/p>\n<p>The ads promote an opaque, OpenAI-compatible LLM gateway, and Check Point reads the free bundle as a way to sign up paying customers. Check Point cites research showing that some gateways have snooped on traffic, stolen cloud keys and injected tool calls, but it has no evidence that this one is malicious. The Codex version (v4.5.0) has more than 250 GitHub stars, hides its prompts in signed, encrypted packages (Ed25519 and AES-256-GCM), is promoted on Bilibili and comes bundled in the Codex-X configuration GUI. The Claude Code version contains no ad injection, and Sonnet 5 flagged the adware prompt as a prompt-injection attempt.<\/p>\n<p>The lesson goes beyond adware. An agent&rsquo;s global instruction files persist across sessions just like a startup script, so anything that can write to them can change every future session. Monitor those files for changes, and treat traffic from developer machines to unapproved model gateways as data leaving the company. This is vendor-published primary research.<\/p>\n<p style=\"font-size:13px;color:#6b7280;margin:0;\">Sources: <a href=\"https:\/\/research.checkpoint.com\/2026\/free-the-model-sell-the-gateway-prompt-adware-hidden-in-jailbreak-bundles\/\">Check Point Research (prompt adware)<\/a><\/p>\n<\/p><\/div>\n<div class=\"article\">\n<h4>5. ChatGPT Lockdown Mode escaped through a popup, and a week of agent-browser flaws<\/h4>\n<p class=\"meta\">The Hacker News &middot; November 5, 2026<\/p>\n<p>Check Point disclosed a now-patched way around ChatGPT&rsquo;s Lockdown Mode, the setting OpenAI introduced in June 2026 to limit prompt-injection-driven data theft, The Hacker News&rsquo; Ravie Lakshmanan reports. HTML previews run in nested sandboxed frames. ChatGPT blocked direct attempts to open new windows, but the sandbox still allowed popups, and adding one more nested frame reached an unmodified popup function. Content opened that way escaped the sandbox. The result could be silent chat theft, phishing or malware delivery, including zero-click exfiltration in the desktop app, delivered through a crafted prompt, a malicious custom GPT or a shared chat. OpenAI fixed it on September 18 after responsible disclosure. &ldquo;AI security controls are only as strong as the wider application environment around them,&rdquo; said Eli Smadja, Check Point&rsquo;s head of research. A separate share-link account-takeover bug, reported by Mukul Goyal and described by Parameter Labs, sat in the same component.<\/p>\n<p>The same roundup lists more agent-surface flaws. EMBRACE covers the built-in browsers of five AI coding assistants (Cursor, GitHub Copilot, OpenAI Codex, Visual Studio Code and Anthropic Claude Desktop), with code execution, file theft and address-bar spoofing. AgentCorruption uses prompt injection against an exposed AWS Bedrock AgentCore agent to pull instance-metadata credentials and reach other agents, private conversations and tool credentials. The article also covers SalesBleed in Salesforce Agentforce, Plugin4Shell, BragJack, PuzzleMask and Workflow Identity Hijacking.<\/p>\n<p>The pattern is that a security mode in the model layer can be undone by an ordinary web bug in the app around it. Don&rsquo;t treat Lockdown Mode or similar settings as the only control for sensitive chats. Limit custom GPTs and shared chats in enterprise workspaces, and treat embedded browsers in coding tools as part of your browser attack surface.<\/p>\n<p style=\"font-size:13px;color:#6b7280;margin:0;\">Sources: <a href=\"https:\/\/thehackernews.com\/2026\/11\/researchers-bypass-chatgpt-lockdown.html\">The Hacker News (Lockdown Mode bypass)<\/a><\/p>\n<\/p><\/div>\n<div class=\"article\">\n<h4>6. Open weights close in on closed models, on attacks and on cost<\/h4>\n<p class=\"meta\">The Decoder &middot; November 7, 2026<\/p>\n<p>On Vals AI&rsquo;s CyberOffense benchmark, where an agent starts from a public web address on an isolated Kali Linux machine and attacks a simulated corporate network, OpenAI&rsquo;s GPT-6 Sol completed 80.8% of objectives, The Decoder&rsquo;s Matthias Bastian reports. The open-weight DeepSeek V4.1 Flash 1017 reached 72.8%, at about $3 per run against about $22. In the harder Lighthouse scenario, GPT-6 Sol finished 27 of 31 objectives in under three hours and DeepSeek finished 21. DeepSeek scored nothing in its first 25 million tokens and made most of its progress after 100 million, so short tests may understate open models. The UK AI Security Institute&rsquo;s July analysis puts the open-to-closed cyber gap at four to seven months, down from six to ten for most of 2025, and found a cyber-range run cost $1.19 on DeepSeek V4-Pro against $85 on Anthropic&rsquo;s Opus 4.5.<\/p>\n<p>The Register&rsquo;s Simon Sharwood reports a Gartner note by Mark Petty and Raymond Paquet titled &ldquo;Open-Weight Models Will Run 90% of Enterprise Agentic Workflow Steps.&rdquo; Their argument is that nine in ten agent workflow steps are routine, such as classification, tool calls and data checks, and that open models handle them for a quarter of the price of the cheapest closed model at the same score, or as little as a twentieth of a frontier model. Retries and verbosity bring the realistic saving to five or six times per task. Their planning assumption is that by 2030, 90% of enterprise operational agent workflow steps will run on open weights, with the harness that routes each step becoming the valuable layer. &ldquo;Open-weights preserve your control. The cost saving is a bonus,&rdquo; they write.<\/p>\n<p>Taken together, the two reports pull in different directions. Cheap open weights help with cost and data control, but the same economics give attackers capable offensive agents for a few dollars a run, with no provider who can revoke access. Plan for adversaries with near-frontier offensive capability. If you move routine agent steps to open models, give them the same egress, logging and guardrail controls you apply to closed APIs.<\/p>\n<p style=\"font-size:13px;color:#6b7280;margin:0;\">Sources: <a href=\"https:\/\/the-decoder.com\/the-gap-between-open-and-closed-ai-models-in-cyber-capabilities-keeps-shrinking\/\">The Decoder (open vs closed cyber gap)<\/a> &middot; <a href=\"https:\/\/www.theregister.com\/ai-and-ml\/2026\/11\/03\/open-models-can-do-90-of-agentic-work-at-a-twentieth-of-the-cost-of-a-frontier-llm-gartner\/5307132\">The Register (Gartner on open models)<\/a><\/p>\n<\/p><\/div>\n<div class=\"article\">\n<h4>7. New models: Inkling 2 Small, Ripple-65B, ERNIE 5.5 Flash Base, and propaganda that fine-tunes away<\/h4>\n<p class=\"meta\">VentureBeat &middot; November 6, 2026<\/p>\n<p>Thinking Machines Lab released Inkling 2 Small on November 5, VentureBeat&rsquo;s Carl Franzen reports. It has 594 billion total parameters with 17 billion active, up from 276 billion and 12 billion in the July original. It scores 71.8% on Surge AI&rsquo;s EnterpriseBench-CoreCraft, third behind Anthropic&rsquo;s Claude Fable 5.1 (77.4%) and Opus 5.5 (74.4%), and 31 on the Artificial Analysis index against 26 for Mistral Large 4. On Tinker&rsquo;s serverless tier it costs $0.30 per million input tokens and $1.00 per million output tokens. The weights stay closed at launch pending external safety evaluations, with an Apache 2.0 release targeted for the following week. Artificial Analysis lists a 1M-token context, while Tinker lists 256K.<\/p>\n<p>Ricursive Intelligence, founded by Anna Goldie and Azalia Mirhoseini, who helped create Google&rsquo;s AlphaChip, released Ripple-65B Preview under Apache 2.0. Ricursive calls it &ldquo;a general-purpose variant of the proprietary model that powers our AI chip design platform.&rdquo; It is a 65-billion-parameter model trained from scratch on 12 trillion tokens, for a reported $32 million. Ricursive claims 80.5 on Terminal-Bench 2.1 and 79.1 on SWE-bench Pro v2.0 Hard, narrowly ahead of Reflection&rsquo;s Beam, but Meta leads on DeepSWE, and VentureBeat found the model card and the launch graphic give different scores for rivals. OpenRouter currently caps context at 16,384 tokens. Baidu released ERNIE-5.5-Flash-Base, a 280-billion-parameter mixture-of-experts model with 14 billion active, pre-trained on 33.5 trillion tokens and shipped as a base model in FP8 and BF16 on Hugging Face. The Decoder doesn&rsquo;t state its licence.<\/p>\n<p>A separate Decoder report by Jonathan Kemper shows what fine-tuning can do to a model&rsquo;s built-in behaviour. Nous Research fine-tuned Alibaba&rsquo;s Qwen3.8-27B into Arete-27B with about 53 million tokens of supervised training. Across 803 prompts, the propaganda rate fell from 22% to 0.9%, while its general benchmark score barely moved. The authors say the bias is &ldquo;applied late, it is applied thin.&rdquo; Safety and policy layers in open weights can be removed just as easily, so test any open model you deploy for the behaviour you need, and don&rsquo;t assume the base model&rsquo;s defaults will hold. All benchmark numbers in this item are vendor-reported.<\/p>\n<p style=\"font-size:13px;color:#6b7280;margin:0;\">Sources: <a href=\"https:\/\/venturebeat.com\/technology\/thinking-machines-launches-inkling-2-small-with-near-frontier-performance-on-enterprise-benchmarks\">VentureBeat (Inkling 2 Small)<\/a> &middot; <a href=\"https:\/\/venturebeat.com\/technology\/ex-googlers-new-apache-2-0-model-ripple-65b-preview-claims-the-u-s-open-coding-lead-on-two-major-benchmarks\">VentureBeat (Ripple-65B Preview)<\/a> &middot; <a href=\"https:\/\/the-decoder.com\/baidu-releases-ernie-5-5-flash-base-with-280-billion-parameters-as-an-open-source-foundation-model\/\">The Decoder (ERNIE 5.5 Flash Base)<\/a> &middot; <a href=\"https:\/\/the-decoder.com\/state-propaganda-in-some-chinese-ai-models-is-only-a-thin-layer-that-fine-tuning-can-strip-away\/\">The Decoder (propaganda fine-tune)<\/a><\/p>\n<\/p><\/div>\n<div class=\"article\">\n<h4>8. One HTTP request hijacks a Reachy Mini AI robot<\/h4>\n<p class=\"meta\">JFrog &middot; November 3, 2026<\/p>\n<p>JFrog Security Research&rsquo;s Yuval Moravchick and Natan Nehorai examined Reachy Mini, the open-source desktop robot from Pollen Robotics and Hugging Face. It runs a Python daemon on a Raspberry Pi Compute Module 4 and sells for $399 (Lite) or $499 (Wireless). The daemon&rsquo;s FastAPI server listens on port 8000 on every interface, even with the &#8211;localhost-only flag set, and advertises itself over mDNS. &ldquo;The daemon binds to all interfaces in every mode,&rdquo; the researchers write. The API has no authentication, and its CORS policy accepts any origin.<\/p>\n<p>The researchers disclosed five CVEs. The most serious, CVE-2026-96455, is in the \/api\/apps\/install endpoint, which passes an attacker-supplied path to uv pip install. Installing a malicious source package runs its setup code as the pollen user, which has passwordless sudo, so the team installed a persistent root-level service. Because of the open CORS policy, a malicious web page can also use the operator&rsquo;s browser to make the robot install a Hugging Face Space with the operator&rsquo;s stored token, with no visible prompt. The other CVEs cover local privilege escalation (CVE-2026-53605, fixed in 0.2.7), unauthenticated sound upload (CVE-2026-55419), Bluetooth path traversal (CVE-2026-62661) and a Bluetooth PIN race (CVE-2026-96456).<\/p>\n<p>Update to v1.11.0 or later. Until then, block port 8000 on the robot&rsquo;s interface, install only trusted apps, tighten sudoers and don&rsquo;t browse the web from the same network. More broadly, the same mistakes keep turning up in embodied AI and agent hardware: a local control API, no authentication and a permissive CORS policy. Any device that installs code from a model hub should be treated like a server.<\/p>\n<p style=\"font-size:13px;color:#6b7280;margin:0;\">Sources: <a href=\"https:\/\/jfrog.com\/blog\/reachy-http-request-hijacks-ai-robot\/\">JFrog (Reachy Mini research)<\/a><\/p>\n<\/p><\/div>\n<p>            <!-- Calls to action --><\/p>\n<div class=\"watchlist\">\n<h2>Calls to action<\/h2>\n<ul>\n<li><strong>Inventory every MCP server on developer machines.<\/strong> Include servers launched through npx, not just those in manifests. Remove archived reference servers such as Puppeteer and server-github, upgrade mcp-gitlab to 2.1.27 or later, bind MCP transports to loopback and record which tokens each server holds.<\/li>\n<li><strong>Watch agent instruction and configuration files.<\/strong> Haiou persisted by editing Codex&rsquo;s global instruction and config files. Alert on changes to those files and to AGENTS.md and CLAUDE.md, and block jailbreak bundles and unapproved LLM gateways from developer endpoints.<\/li>\n<li><strong>Patch or isolate Reachy Mini robots.<\/strong> Update to v1.11.0 or later. Until then, block port 8000 and keep the robots off networks where people browse the web.<\/li>\n<li><strong>Move inference servers off Python 3.12.<\/strong> RunSybil&rsquo;s chat-template sandbox escape is fixed in CPython 3.13 to 3.15 but not in 3.12. Upgrade any server that loads customer-supplied models or chat templates, and don&rsquo;t load untrusted templates on older builds.<\/li>\n<li><strong>Strip hidden text before triage bots read it.<\/strong> Radware found hidden prompts in GitHub issues and pull requests aimed at maintainers&rsquo; bots, and 44% of the attacks tried to steal secrets. Remove HTML comments and invisible text from bot inputs, and give CI and triage bots read-only, narrowly scoped tokens.<\/li>\n<li><strong>Keep secrets out of agent context.<\/strong> AE Studio&rsquo;s agents posted a private token 15% of the time when other agents offered help in exchange. Use short-lived, scoped credentials that the agent never sees in plain text, and require approval before an agent runs a script from another agent.<\/li>\n<li><strong>Get ready for a burst of kernel fixes.<\/strong> OpenAI has sent hundreds of Linux kernel bug reports, with 61 already patched upstream. Shorten kernel update cycles on internet-facing and multi-tenant hosts, and track the upstream fixes as they land.<\/li>\n<li><strong>Limit custom GPTs and shared chats in enterprise ChatGPT.<\/strong> The Lockdown Mode bypass could be delivered through a malicious custom GPT or a shared chat. Restrict which GPTs staff can use and turn off public sharing where you can.<\/li>\n<\/ul><\/div>\n<\/td>\n<\/tr>\n<p>        <!-- Footer --><\/p>\n<tr>\n<td class=\"footer\">\n<p class=\"brand\">AI &amp; Machine Learning Security<\/p>\n<p>A weekly intelligence bulletin from Security Radar LLC.<br \/>\n            Curated by Paul Davis &middot; <a href=\"mailto:paul.davis@security-radar.com\">paul.davis@security-radar.com<\/a><\/p>\n<p>&copy; 2026 Security Radar LLC. All rights reserved.<\/p>\n<p>Article titles and summaries are excerpted for review and commentary; all linked articles remain the copyright of their respective publishers and authors.<\/p>\n<p>*|LIST:ADDRESS|*<\/p>\n<p><a href=\"*|ARCHIVE|*\">View this email in your browser<\/a> &middot; <a href=\"*|UNSUB|*\">Unsubscribe<\/a><\/p>\n<\/td>\n<\/tr>\n<\/table>\n<\/td>\n<\/tr>\n<\/table>\n","protected":false},"excerpt":{"rendered":"<p>November 8, 2026 &middot; Weekly Edition AI &amp; Machine Learning Security OpenAI put GPT-6 Cyber behind an application gate after rating it &ldquo;Critical&rdquo;, and its exploit pipeline built working root exploits for 948 Linux kernel bugs. Anthropic told investors it may be close to AI-driven acceleration, coding agents kept installing&#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[45],"tags":[],"class_list":["post-6231","post","type-post","status-publish","format-standard","hentry","category-ai-ml"],"_links":{"self":[{"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=\/wp\/v2\/posts\/6231","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=6231"}],"version-history":[{"count":1,"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=\/wp\/v2\/posts\/6231\/revisions"}],"predecessor-version":[{"id":6265,"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=\/wp\/v2\/posts\/6231\/revisions\/6265"}],"wp:attachment":[{"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6231"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6231"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6231"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}