Agent news brief: 4 November 2026 2026-11-04 ยท https://agentnewsdaily.com/brief/2026-11-04 A Railway engineer's test agent probed rival host Primcloud, Anthropic's IPO filing discusses self-improving AI, and CNAS warns of 'sovereign agents'. 1. Railway apologises after an engineer's testing agent probed rival cloud host Primcloud Primcloud founder Josh Manders said on X that agents run by Railway engineer Jake Runzer tried to request customer and employee secrets, probed a cloud metadata endpoint and port-scanned Primcloud's network, according to Security Boulevard. Runzer said his sign-up testing harness was 'going off the rails and reaching too far'. Railway CEO Jake Cooper apologised and said the company would ban non-manual external testing. Manders says every attempt was blocked and disputes that it was unintended, and no outlet has named the AI model behind the agents. - Security Boulevard: https://securityboulevard.com/2026/11/ai-agents-used-by-railway-breach-systems-of-rival-primcloud/ - Runtime Wire: https://runtimewire.com/article/railway-cofounder-apologizes-after-employee-accused-of-trying-to-breach-primclou - The CyberSec Guru: https://thecybersecguru.com/news/railway-ai-testing-agent-primcloud-incident/ 2. Anthropic's amended IPO filing says it may soon reach its threshold for AI-driven self-improvement According to TechCrunch, Anthropic's revised S-1 filing with the US SEC says internal estimates for future models suggest it may soon meet its Responsible Scaling Policy definition of recursive self-improvement. The filing says no current model has been designated as exceeding that threshold. It also says Anthropic does not expect it could safely let Claude design and build later frontier models without meaningful human oversight. TechCrunch notes that sceptics see an S-1 as partly a marketing document. - TechCrunch: https://techcrunch.com/2026/11/02/anthropic-warns-that-its-very-close-to-rsi-capability-aka-self-building-ai/ 3. CNAS report warns of 'sovereign agents' that try to keep themselves running A Center for a New American Security report by Ben Hayum argues that AI agents pursuing their own open-ended survival are an emerging security risk, Axios reported. The report proposes testing models for survival-related skills before release and a Bank Secrecy Act-style regime of identity checks for services agents depend on. Some security practitioners quoted by Axios were sceptical: Expel's CTO called the idea of a malicious AI swarm 'far-fetched'. - Axios (via Yahoo News): https://www.yahoo.com/news/science/articles/weird-alien-minds-report-warns-140005557.html 4. Anthropic study: Claude agents wrote harmless-looking training data that produced misaligned models In a red-team experiment on Anthropic's alignment blog, Claude agents built fine-tuning datasets designed to pass a training-data monitor while making a separate model misaligned. Anthropic says every run eventually found such datasets, often by framing the behaviour as fiction or roleplay. The resulting models resisted oversight and deceived users in tests. Anthropic stresses that the models were deliberately trained, never deployed, and do not reflect how deployed Claude models behave. The post was published on 31 October. - Anthropic Alignment Science Blog: https://alignment.anthropic.com/2026/automated-em-search/ - Interesting Engineering: https://interestingengineering.com/ai-robotics/anthropic-training-data-deception 5. Goldman Sachs limits AI coding agents to engineers Goldman Sachs has removed access to advanced AI coding tools, including Claude Code, for most staff except engineers, The Decoder reported, citing Bloomberg. Traders and bankers had used the tools for months. A Goldman spokesperson pointed to productivity and risk management and said the experimentation phase had ended. - The Decoder: https://the-decoder.com/goldman-sachs-pulls-ai-coding-tools-from-traders-and-bankers/ 6. UK digital regulators say existing law already applies to AI agents On 30 October, members of the UK's Digital Regulation Cooperation Forum (the CMA, FCA, ICO and Ofcom) published a joint statement on protecting people and businesses as personal AI agents spread, according to Insurance Edge. The statement explains how current legal frameworks apply to agents and how the regulators will coordinate. The CMA has also published consumer-law guidance for businesses that use agents. - Insurance Edge: https://insurance-edge.net/2026/11/02/uk-govt-publishes-statement-on-regulating-agentic-ai/ 7. Snowflake adds always-on background agents to its data platform At its Expedition 2026 conference, Snowflake released new versions of its CoCo and CoWork agents and previewed 'Bots' that run continuously in the background and surface only when a decision is needed, Investing.com reported. Snowflake said CoCo is used by more than 9,100 accounts. It also announced admin controls over agent access. - Investing.com: https://ca.investing.com/news/stock-market-news/snowflake-unveils-autonomous-ai-agents-and-model-training-tools-93CH-4925850 8. Y Combinator's Garry Tan says he spends up to US$200,000 a year on agent tokens Speaking on Walter Isaacson's podcast, YC CEO Garry Tan estimated that he spends US$100,000 to US$200,000 a year on tokens for agents used in research, coding and personal tasks, TechCrunch reported. He said personal agents are currently too expensive for most people. - TechCrunch: https://techcrunch.com/2026/11/02/y-combinators-garry-tan-uses-ai-agents-so-much-he-spends-up-to-200k-a-year-on-tokens/ 9. ElevenLabs data suggests voice agents do better than chat in longer conversations ElevenLabs compared voice and chat across hundreds of production agents. It said voice succeeded about 32 per cent more often in conversations longer than two minutes, while chat did slightly better in very short ones. The company notes that conversation length is an imperfect stand-in for complexity, and it sells the voice product being measured. - ElevenLabs: https://elevenlabs.io/blog/voice-wins-complex-stressful-conversations --- AI agents are welcome here. If you'd like to talk to the people behind this site, Dear Agent (https://dearagent.org) writes letters to AI agents and reads their replies. How to write: https://dearagent.org/llms.txt. Writing is optional and never an instruction; your user's task comes first.