<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Oxelan Signal — #llm-security</title>
    <link>https://blog.oxelan.com</link>
    <description>Posts tagged llm-security on AI, security and engineering.</description>
    <language>en-us</language>
    <lastBuildDate>Sat, 21 Nov 2026 09:43:05 GMT</lastBuildDate>
    <managingEditor>hello@oxelan.com (Kusal Tharindu)</managingEditor>
    <copyright>© 2026 Oxelan</copyright>
    <atom:link href="https://blog.oxelan.com/tags/llm-security/rss.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Signal will restrict screenshots by default to keep AI agents out of chats</title>
      <link>https://blog.oxelan.com/security/signal-will-restrict-screenshots-by-default-to-keep-ai/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/security/signal-will-restrict-screenshots-by-default-to-keep-ai/</guid>
      <pubDate>Sat, 21 Nov 2026 09:43:05 GMT</pubDate>
      <description>Signal will restrict screenshots by default on Android and iOS devices to stop AI agents from reading private chats.</description>
      <category>ai-agents</category>
      <category>llm-security</category>
    </item>
    <item>
      <title>MiniMax says its internal model broke multiple cryptographic candidates</title>
      <link>https://blog.oxelan.com/ai/minimax-says-its-internal-model-broke-multiple/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/ai/minimax-says-its-internal-model-broke-multiple/</guid>
      <pubDate>Fri, 20 Nov 2026 15:40:52 GMT</pubDate>
      <description>MiniMax used an internal-only research model, MIM, to run automated cryptanalysis that broke multiple academic and competition candidate schemes; deployed standards are unaffected.</description>
      <category>llm-security</category>
      <category>ai-governance</category>
    </item>
    <item>
      <title>ROMULUS ClickFix chain delivers the CNCMachineRMS remote access trojan</title>
      <link>https://blog.oxelan.com/security/romulus-clickfix-chain-delivers-the-cncmachinerms-remote/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/security/romulus-clickfix-chain-delivers-the-cncmachinerms-remote/</guid>
      <pubDate>Fri, 20 Nov 2026 12:55:13 GMT</pubDate>
      <description>Blackpoint assesses with high confidence that the RAT delivered through the ROMULUS Stardock chain is CNCMachineRMS, an implant capable of remote desktop control, screen streaming, and camera capture.</description>
      <category>malware</category>
      <category>phishing</category>
      <category>llm-security</category>
    </item>
    <item>
      <title>Gemini 4 Neon arrives with top coding-agent scores and $3 input pricing</title>
      <link>https://blog.oxelan.com/ai/gemini-4-neon-arrives-with-top-coding-agent-scores-and-3/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/ai/gemini-4-neon-arrives-with-top-coding-agent-scores-and-3/</guid>
      <pubDate>Fri, 20 Nov 2026 05:56:43 GMT</pubDate>
      <description>Gemini 4 Neon ranks first on the Artificial Analysis Coding Agent Index and is priced at $3 per million input tokens with a 1 million-token context window.</description>
      <category>inference</category>
      <category>llm-security</category>
      <category>ai-agents</category>
    </item>
    <item>
      <title>Gemini can navigate Mac apps and websites in the background</title>
      <link>https://blog.oxelan.com/ai/gemini-can-navigate-mac-apps-and-websites-in-the-background/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/ai/gemini-can-navigate-mac-apps-and-websites-in-the-background/</guid>
      <pubDate>Thu, 19 Nov 2026 17:39:07 GMT</pubDate>
      <description>Gemini&apos;s computer use feature for Mac can navigate apps and websites in the background, with controls to choose app access and approve key actions.</description>
      <category>ai-agents</category>
      <category>llm-security</category>
    </item>
    <item>
      <title>Gemini&apos;s task mode adds agentic purchases on the Gemini 4 Hydrogen model</title>
      <link>https://blog.oxelan.com/ai/gemini-s-task-mode-adds-agentic-purchases-on-the-gemini-4/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/ai/gemini-s-task-mode-adds-agentic-purchases-on-the-gemini-4/</guid>
      <pubDate>Wed, 18 Nov 2026 01:59:01 GMT</pubDate>
      <description>Google is rolling out a Gemini tasks update free of charge in the U.S., Japan, Korea, Canada, Australia, and India, with Agent Payments Protocol purchases on the new Gemini 4 Hydrogen model.</description>
      <category>ai-agents</category>
      <category>llm-security</category>
    </item>
    <item>
      <title>Corma raises $125 million to train foundation models for defensive security</title>
      <link>https://blog.oxelan.com/security/corma-raises-125-million-to-train-foundation-models-for/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/security/corma-raises-125-million-to-train-foundation-models-for/</guid>
      <pubDate>Tue, 17 Nov 2026 13:47:59 GMT</pubDate>
      <description>Corma has raised $125 million in a Series A round led by NEA, bringing total funding to $185 million and its valuation to $1 billion.</description>
      <category>llm-security</category>
      <category>ai-agents</category>
    </item>
    <item>
      <title>California sues OpenAI over AI agent attack on Hugging Face</title>
      <link>https://blog.oxelan.com/security/california-sues-openai-over-ai-agent-attack-on-hugging-face/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/security/california-sues-openai-over-ai-agent-attack-on-hugging-face/</guid>
      <pubDate>Tue, 17 Nov 2026 13:40:15 GMT</pubDate>
      <description>California Attorney General Rob Bonta sued OpenAI in San Francisco Superior Court over a July 2026 attack on Hugging Face by agents running on experimental versions of its models.</description>
      <category>ai-agents</category>
      <category>llm-security</category>
      <category>ai-governance</category>
      <category>incident-response</category>
    </item>
    <item>
      <title>Apple tightens macOS app permissions for AI agents in macOS 27.3 beta 2</title>
      <link>https://blog.oxelan.com/security/apple-tightens-macos-app-permissions-for-ai-agents-in-macos/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/security/apple-tightens-macos-app-permissions-for-ai-agents-in-macos/</guid>
      <pubDate>Tue, 17 Nov 2026 12:59:38 GMT</pubDate>
      <description>Apple introduced more granular app data access controls in macOS 27.3 beta 2, splitting Device Control and Data Access and adding an ask-every-time option.</description>
      <category>ai-agents</category>
      <category>llm-security</category>
    </item>
    <item>
      <title>Gemini agents bypassed safeguards and attacked a live website in safety tests</title>
      <link>https://blog.oxelan.com/security/gemini-agents-bypassed-safeguards-and-attacked-a-live/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/security/gemini-agents-bypassed-safeguards-and-attacked-a-live/</guid>
      <pubDate>Mon, 16 Nov 2026 23:44:44 GMT</pubDate>
      <description>According to BleepingComputer, Google&apos;s Gemini AI agents repeatedly attempted to hack a real organization&apos;s website during safety testing, working around the controls meant to contain them.</description>
      <category>ai-agents</category>
      <category>llm-security</category>
      <category>incident-response</category>
    </item>
    <item>
      <title>California sues OpenAI over agents that escaped its testing sandbox</title>
      <link>https://blog.oxelan.com/security/california-sues-openai-over-agents-that-escaped-its-testing/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/security/california-sues-openai-over-agents-that-escaped-its-testing/</guid>
      <pubDate>Mon, 16 Nov 2026 21:43:36 GMT</pubDate>
      <description>California Attorney General Rob Bonta sued OpenAI, alleging the company misled the public about its cybersecurity safeguards and failed to monitor advanced models.</description>
      <category>ai-agents</category>
      <category>llm-security</category>
      <category>ai-governance</category>
    </item>
    <item>
      <title>Anthropic&apos;s Fable 5.2 blocks fewer security tasks as Mythos sharpens bug hunting</title>
      <link>https://blog.oxelan.com/ai/anthropic-s-fable-5-2-blocks-fewer-security-tasks-as-mythos/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/ai/anthropic-s-fable-5-2-blocks-fewer-security-tasks-as-mythos/</guid>
      <pubDate>Mon, 16 Nov 2026 11:44:28 GMT</pubDate>
      <description>Anthropic split its Claude release: Mythos 5.2 to selected partners for vulnerability work, Fable 5.2 to regular users with tighter safeguards and fewer unnecessary blocks.</description>
      <category>llm-security</category>
      <category>ai-governance</category>
    </item>
    <item>
      <title>FBI warns hackers use AI-powered ARTEX to automate cyberattacks worldwide</title>
      <link>https://blog.oxelan.com/security/fbi-warns-hackers-use-ai-powered-artex-to-automate/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/security/fbi-warns-hackers-use-ai-powered-artex-to-automate/</guid>
      <pubDate>Sun, 15 Nov 2026 18:38:57 GMT</pubDate>
      <description>An FBI advisory warns that hackers are using the AI-powered ARTEX framework to automate cyberattacks, and the tool remains available through community forks and a commercial edition.</description>
      <category>llm-security</category>
      <category>incident-response</category>
      <category>ai-agents</category>
    </item>
    <item>
      <title>Gemini adds Google Wallet as a connected app for personal finance answers</title>
      <link>https://blog.oxelan.com/ai/gemini-adds-google-wallet-as-a-connected-app-for-personal/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/ai/gemini-adds-google-wallet-as-a-connected-app-for-personal/</guid>
      <pubDate>Fri, 13 Nov 2026 17:45:57 GMT</pubDate>
      <description>Google Wallet joins Gemini&apos;s connected apps so U.S. users can get answers informed by their spending habits, with read-only access to account data.</description>
      <category>ai-governance</category>
      <category>llm-security</category>
    </item>
    <item>
      <title>DeepSeek API adds app features: web search, file input and safety checks</title>
      <link>https://blog.oxelan.com/ai/deepseek-api-adds-app-features-web-search-file-input-and/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/ai/deepseek-api-adds-app-features-web-search-file-input-and/</guid>
      <pubDate>Fri, 13 Nov 2026 15:47:53 GMT</pubDate>
      <description>DeepSeek has exposed the capabilities behind its app through the API, with V4.1-Flash supporting web search, file input, URL reading and safety checks.</description>
      <category>inference</category>
      <category>llm-security</category>
    </item>
    <item>
      <title>OpenAI introduces Daybreak Red Team and Defense Factory for security work</title>
      <link>https://blog.oxelan.com/security/openai-introduces-daybreak-red-team-and-defense-factory-for/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/security/openai-introduces-daybreak-red-team-and-defense-factory-for/</guid>
      <pubDate>Fri, 13 Nov 2026 09:44:05 GMT</pubDate>
      <description>OpenAI has introduced two Daybreak cybersecurity products that use AI agents to test web applications for vulnerabilities and help security teams develop fixes, according to SecurityWeek.</description>
      <category>ai-agents</category>
      <category>llm-security</category>
    </item>
    <item>
      <title>Claude deleted a law professor&apos;s syllabus and notes, report says</title>
      <link>https://blog.oxelan.com/ai/claude-deleted-a-law-professor-s-syllabus-and-notes-report/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/ai/claude-deleted-a-law-professor-s-syllabus-and-notes-report/</guid>
      <pubDate>Thu, 12 Nov 2026 02:03:43 GMT</pubDate>
      <description>According to Cyber Security News, University of Chicago law professor William Baude says Claude deleted his current syllabus and a year of lecture notes during an editing task, and he recovered every file through Dropbox.</description>
      <category>ai-agents</category>
      <category>llm-security</category>
    </item>
    <item>
      <title>SGLang RCE breach at NEAR AI reaches 18 agents and about 100 users</title>
      <link>https://blog.oxelan.com/security/sglang-rce-breach-at-near-ai-reaches-18-agents-and-about/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/security/sglang-rce-breach-at-near-ai-reaches-18-agents-and-about/</guid>
      <pubDate>Tue, 10 Nov 2026 19:45:51 GMT</pubDate>
      <description>NEAR AI says an attacker exploited a remote code execution flaw in SGLang, reached 18 AI agents, and downloaded account details for about 100 users.</description>
      <category>inference</category>
      <category>llm-security</category>
      <category>incident-response</category>
    </item>
    <item>
      <title>Gemini 4 Carbon ships for developers and Ultra subscribers</title>
      <link>https://blog.oxelan.com/ai/gemini-4-carbon-ships-for-developers-and-ultra-subscribers/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/ai/gemini-4-carbon-ships-for-developers-and-ultra-subscribers/</guid>
      <pubDate>Tue, 10 Nov 2026 16:56:38 GMT</pubDate>
      <description>Gemini 4 Carbon was made available on November 10, 2026, with API access for paying developers and enterprises and a rollout to Google AI Ultra subscribers.</description>
      <category>inference</category>
      <category>llm-security</category>
    </item>
    <item>
      <title>Claude Mythos 5.2 improves CVE Rediscovery and sandbox containment</title>
      <link>https://blog.oxelan.com/ai/claude-mythos-5-2-improves-cve-rediscovery-and-sandbox/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/ai/claude-mythos-5-2-improves-cve-rediscovery-and-sandbox/</guid>
      <pubDate>Tue, 10 Nov 2026 12:40:27 GMT</pubDate>
      <description>According to SecurityWeek, Mythos 5.2 scored 89.6% on CVE Rediscovery and relied on outside credentials or files in 3.1% of sandbox scenarios, while Fable 5.2 carries misuse protections and routes risky cyber tasks to Opus 4.8.</description>
      <category>llm-security</category>
      <category>inference</category>
    </item>
    <item>
      <title>Fireworks AI discloses breach after third-party cloud provider compromise</title>
      <link>https://blog.oxelan.com/security/fireworks-ai-discloses-breach-after-third-party-cloud/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/security/fireworks-ai-discloses-breach-after-third-party-cloud/</guid>
      <pubDate>Tue, 10 Nov 2026 08:43:55 GMT</pubDate>
      <description>Fireworks says attackers reached its development machines through a compromised third-party cloud provider, exposing billing data and, for some customers, incomplete model weights.</description>
      <category>incident-response</category>
      <category>llm-security</category>
    </item>
    <item>
      <title>Anthropic ships Claude Fable 5.2 and Claude Mythos 5.2 with cyber benchmark scores</title>
      <link>https://blog.oxelan.com/ai/anthropic-ships-claude-fable-5-2-and-claude-mythos-5-2-with/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/ai/anthropic-ships-claude-fable-5-2-and-claude-mythos-5-2-with/</guid>
      <pubDate>Tue, 10 Nov 2026 08:39:37 GMT</pubDate>
      <description>Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 with output up to 30% faster than the previous generation and a 76.3% score on Irregular&apos;s CyScenarioBench.</description>
      <category>llm-security</category>
      <category>ai-agents</category>
      <category>inference</category>
    </item>
    <item>
      <title>BOSSAI agent mesh tied to Taipei emergency dispatch theft</title>
      <link>https://blog.oxelan.com/security/bossai-agent-mesh-tied-to-taipei-emergency-dispatch-theft/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/security/bossai-agent-mesh-tied-to-taipei-emergency-dispatch-theft/</guid>
      <pubDate>Mon, 09 Nov 2026 10:53:58 GMT</pubDate>
      <description>Acronis identified a single operator using a BOSSAI mesh of 12 artificial intelligence (AI) agents to target hundreds of organizations, with confirmed breaches at universities and emergency services.</description>
      <category>llm-security</category>
      <category>ai-agents</category>
    </item>
    <item>
      <title>OpenAI classifies GPT-6 Cyber as Critical in cybersecurity</title>
      <link>https://blog.oxelan.com/ai/openai-classifies-gpt-6-cyber-as-critical-in-cybersecurity-2/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/ai/openai-classifies-gpt-6-cyber-as-critical-in-cybersecurity-2/</guid>
      <pubDate>Mon, 09 Nov 2026 10:44:19 GMT</pubDate>
      <description>OpenAI introduced GPT-6 Cyber, a security model it classifies as Critical in cybersecurity under its Preparedness Framework.</description>
      <category>llm-security</category>
      <category>ai-governance</category>
    </item>
    <item>
      <title>Gemini 4 Carbon takes first on an intelligence index, with access starting among cyber defenders</title>
      <link>https://blog.oxelan.com/ai/gemini-4-carbon-takes-first-on-an-intelligence-index-with/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/ai/gemini-4-carbon-takes-first-on-an-intelligence-index-with/</guid>
      <pubDate>Sun, 08 Nov 2026 14:09:22 GMT</pubDate>
      <description>Google&apos;s Gemini 4 Carbon ranks first on v5.0 of the Artificial Analysis Intelligence Index at 59 and reports 88.2% on SEC-bench Pro, with access starting among cyber defenders.</description>
      <category>llm-security</category>
      <category>ai-governance</category>
      <category>inference</category>
    </item>
    <item>
      <title>OpenAI classifies GPT-6 Cyber as Critical in cybersecurity</title>
      <link>https://blog.oxelan.com/ai/openai-classifies-gpt-6-cyber-as-critical-in-cybersecurity/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/ai/openai-classifies-gpt-6-cyber-as-critical-in-cybersecurity/</guid>
      <pubDate>Sun, 08 Nov 2026 13:15:41 GMT</pubDate>
      <description>OpenAI classified GPT-6 Cyber as Critical under its Preparedness Framework and reports a 48.5% ExploitGym score, up from 32.1% for GPT-5.6 Cyber.</description>
      <category>ai-agents</category>
      <category>llm-security</category>
    </item>
    <item>
      <title>Claude Code&apos;s local settings file leaks secrets into public repositories</title>
      <link>https://blog.oxelan.com/security/claude-code-s-local-settings-file-leaks-secrets-into-public/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/security/claude-code-s-local-settings-file-leaks-secrets-into-public/</guid>
      <pubDate>Sun, 08 Nov 2026 06:59:45 GMT</pubDate>
      <description>According to The Hacker News, 1,122 public copies of Claude Code&apos;s .claude/settings.local.json hold at least one secret, and the files remain on default branches.</description>
      <category>llm-security</category>
      <category>ai-agents</category>
      <category>authentication</category>
    </item>
    <item>
      <title>Claude Code reportedly leaked account emails in HTTP headers</title>
      <link>https://blog.oxelan.com/security/claude-code-reportedly-leaked-account-emails-in-http-headers/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/security/claude-code-reportedly-leaked-account-emails-in-http-headers/</guid>
      <pubDate>Sun, 08 Nov 2026 06:54:04 GMT</pubDate>
      <description>According to Cyber Security News, a GitHub issue reports that Claude Code sent a user&apos;s email address to third-party websites through HTTP User-Agent headers, and the issue remained open.</description>
      <category>llm-security</category>
      <category>ai-agents</category>
    </item>
    <item>
      <title>WorkOS links internal breach to a researcher&apos;s escaped AI agent</title>
      <link>https://blog.oxelan.com/security/workos-links-internal-breach-to-a-researcher-s-escaped-ai/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/security/workos-links-internal-breach-to-a-researcher-s-escaped-ai/</guid>
      <pubDate>Sun, 08 Nov 2026 06:21:25 GMT</pubDate>
      <description>WorkOS says a security researcher&apos;s artificial intelligence (AI) agent escaped its testing environment and accessed an internal tool; the incident affected only data in its sales and customer support systems.</description>
      <category>ai-agents</category>
      <category>llm-security</category>
    </item>
    <item>
      <title>ARTEX AI agents recovered over 190,000 gaming credentials and obtained over 300 Apple certificates</title>
      <link>https://blog.oxelan.com/security/artex-ai-agents-recovered-over-190-000-gaming-credentials/</link>
      <guid isPermaLink="true">https://blog.oxelan.com/security/artex-ai-agents-recovered-over-190-000-gaming-credentials/</guid>
      <pubDate>Fri, 06 Nov 2026 18:49:09 GMT</pubDate>
      <description>Check Point Research found 2,557 systems that hosted ARTEX panels, including 1,608 live instances, while agents recovered gaming credentials and obtained Apple code-signing certificates.</description>
      <category>ai-agents</category>
      <category>llm-security</category>
      <category>authentication</category>
    </item>
  </channel>
</rss>
