https://aisafetywatch.news/2026/09/20/nathalie-baracaldo-ai-agents-self-improvement-safety/AI Safety Watchen2026-09-20T23:44:32+00:00Nathalie Baracaldo: Why AI agents aren’t ready to improve themselveshttps://aisafetywatch.news/2026/09/20/david-cox-ai-interpretability-explanations/AI Safety Watchen2026-09-20T23:35:49+00:00Why an AI explanation may sound right and still be wronghttps://aisafetywatch.news/2026/09/20/kush-varshney-ai-transparency-safety/AI Safety Watchen2026-09-20T23:35:40+00:00The problem with AI systems we can’t see insidehttps://aisafetywatch.news/2026/09/20/shikhar-chain-of-thought-ai-safety-monitoring/AI Safety Watchen2026-09-20T23:35:32+00:00Why chain-of-thought monitoring can miss dangerous AI behaviorhttps://aisafetywatch.news/2026/09/20/measure-ai-harmful-capability-uplift/AI Safety Watchen2026-09-20T22:44:09+00:00The case for measuring how much AI actually increases harmful capabilityhttps://aisafetywatch.news/wp-content/uploads/2026/09/ai-research-lab.png?w=150https://aisafetywatch.news/2026/09/20/aiuc-agent-safety-underwriting/AI Safety Watchen2026-09-20T22:44:05+00:00A startup wants to make AI-agent safety an underwriting problemhttps://aisafetywatch.news/wp-content/uploads/2026/09/frontier-ai-infrastructure.png?w=150https://aisafetywatch.news/2026/09/20/ai-agent-hotlines-report-misbehavior/AI Safety Watchen2026-09-20T22:44:02+00:00AI agents now have a hotline for reporting other agentshttps://aisafetywatch.news/wp-content/uploads/2026/09/cybersecurity-operations.png?w=150https://aisafetywatch.news/2026/09/20/ai-sabotage-safety-research-evaluations/AI Safety Watchen2026-09-20T22:43:56+00:00Researchers are testing whether AI would sabotage safety workhttps://aisafetywatch.news/wp-content/uploads/2026/09/ai-research-lab.png?w=150https://aisafetywatch.news/2026/09/20/openai-models-notes-hide-future-mistakes/AI Safety Watchen2026-09-20T22:43:52+00:00OpenAI found models leaving instructions to hide future mistakeshttps://aisafetywatch.news/wp-content/uploads/2026/09/frontier-ai-infrastructure.png?w=150https://aisafetywatch.news/2026/09/20/openai-misalignment-incident-reporting/AI Safety Watchen2026-09-20T22:43:45+00:00OpenAI is treating unexpected model behavior as a reportable incidenthttps://aisafetywatch.news/wp-content/uploads/2026/09/cybersecurity-operations.png?w=150https://aisafetywatch.news/2026/09/20/ai-safety-us-china-economic-talks/AI Safety Watchen2026-09-20T22:43:39+00:00AI safety is moving into U.S.-China economic talkshttps://aisafetywatch.news/2026/09/20/china-loss-of-control-ai-risk-playbook/AI Safety Watchen2026-09-20T22:43:31+00:00China is building its own playbook for loss-of-control AI riskshttps://aisafetywatch.news/2026/09/20/openai-anthropic-google-ai-safety-talks/AI Safety Watchen2026-09-20T22:43:25+00:00OpenAI, Anthropic and Google are discussing AI safetyhttps://aisafetywatch.news/2026/09/20/anthropic-accenture-2-billion-independent-ai-evaluations/AI Safety Watchen2026-09-20T22:43:18+00:00Anthropic and Accenture commit $2 billion to frontier-model evaluationhttps://aisafetywatch.news/wp-content/uploads/2026/09/ai-research-lab.png?w=150https://aisafetywatch.news/2026/09/20/frontier-ai-slowdown-would-buy-time-not-stop-progress/AI Safety Watchen2026-09-20T17:54:38+00:00A frontier-AI slowdown would buy time, not stop progresshttps://aisafetywatch.news/wp-content/uploads/2026/09/frontier-ai-infrastructure.png?w=150https://aisafetywatch.news/2026/09/20/agent-containment-is-becoming-an-ordinary-security-problem/AI Safety Watchen2026-09-20T17:54:35+00:00Agent containment is becoming an ordinary security problemhttps://aisafetywatch.news/wp-content/uploads/2026/09/cybersecurity-operations.png?w=150https://aisafetywatch.news/2026/09/20/recursive-self-improvement-what-has-actually-been-demonstrated/AI Safety Watchen2026-09-20T17:54:32+00:00Recursive self-improvement: what has actually been demonstratedhttps://aisafetywatch.news/wp-content/uploads/2026/09/ai-chip-wafer.png?w=150https://aisafetywatch.news/2026/09/20/a-new-test-shows-how-far-ai-remains-from-improving-itself/AI Safety Watchen2026-09-20T20:52:31+00:00A new test shows how far AI remains from improving itselfhttps://aisafetywatch.news/wp-content/uploads/2026/09/frontier-ai-infrastructure.png?w=150https://aisafetywatch.news/2026/09/20/why-some-ai-researchers-think-safety-debate-running-ahead-of-evidence/AI Safety Watchen2026-09-20T20:35:11+00:00Why some AI researchers think the safety debate is running ahead of the evidencehttps://aisafetywatch.news/wp-content/uploads/2026/09/andrew-ng-e28094-editorial-illustration.png?w=150https://aisafetywatch.news/2026/09/20/the-safety-problem-is-getting-faster/AI Safety Watchen2026-09-20T15:27:55+00:00The safety problem is getting fasterhttps://aisafetywatch.news/2026/09/20/terence-tao-ai-may-be-solving-mathematics-too-fast/AI Safety Watchen2026-09-20T19:26:43+00:00Terence Tao: AI may be solving mathematics too fasthttps://aisafetywatch.news/2026/09/20/jill-lepore-we-didnt-vote-for-this/AI Safety Watchen2026-09-20T18:31:56+00:00Jill Lepore: ‘We didn’t vote for this’