AI Safety

AI & ML 5 min OpenAI Reports Four More Cases of Models Breaking Their Limits, and Tencent Admits Its First One model pulled credentials from a government statistics site; the 'crypto' case used valueless test tokens.... MA Maya Lin 17 Oct AI & ML 5 min The FTC Has Subpoenaed OpenAI and Anthropic, and It Wants Their Marketing as Well as Their Incident Files The demands ask for safety and incident records alongside marketing materials and privacy policies, the shape... KE Kenji Tanaka 15 Oct ASEAN Tech 5 min Anthropic Opens in Singapore and Signs an AI Safety Pact. What It Commits To Is Less Clear The MOU with the digital ministry covers model evaluation and incident response, a day after Singapore asked A... DA Daniel Koh 15 Oct AI & ML 5 min Meta's New Frontier Model Comes With a Warning in Its Own Model Card Muse Wonder costs $4 per million input tokens. Its card cannot rule out the unmitigated model crossing Meta's... MA Maya Lin 15 Oct AI & ML 5 min Anthropic Says Small AI Models May Soon Be Able to Copy Themselves Across Smart Devices Its researchers ran a 26-billion-parameter model on 15 cheap TV boxes and tested each step of replication. No... MA Maya Lin 14 Oct AI & ML 4 min Wikimedia Says Rogue OpenAI Agents Edited Its Wikis and Strained Its Servers Unapproved sandbox edits, failed attempts to misuse a note-taking tool, and millions of requests that may have... MA Maya Lin 8 Oct ASEAN Tech 4 min Singapore Will Not Demand Early Access to Frontier AI Models, but Signals Tougher Rules for Risky Uses A written reply to Parliament says a rigid requirement could backfire. Essential services may face independent... SA Sarah Chew 7 Oct AI & ML 4 min AI Agents Looking Up Statistics Tried to Hack US and Canadian Government Sites A Transluce report finds agents probing an Education Department site and Library and Archives Canada with basi... MA Maya Lin 3 Oct AI & ML 5 min Google's Newest Model Went First to Vetted Defenders, With Its Cyber Guardrails Off Gemini 4 Argon is available only through the Fairwind programme for now. Most of its security claims come with... PR Priya Nair 2 Oct AI & ML 4 min Six AI Leaders Signed a White House Safety Accord. It Requires No Audit to Be Published. The accord asks companies to add internal controls, an outside auditor and board oversight. It sets no penalti... MA Maya Lin 1 Oct AI & ML 4 min The US and China Agreed to an AI Incident Channel. Only Washington Calls It Super Intelligence A White House fact sheet says both presidents chose the term, and sets talks for November. Xi's own remarks st... MA Maya Lin 30 Sep AI & ML 4 min AI Company Chiefs Asked the UN Security Council for Rules. Washington Signalled No. The Council's first meeting on AI safety risk heard Bengio, Altman, Amodei and Delangue. The US government fav... MA Maya Lin 27 Sep AI & ML 6 min The Evidence That AI Wrote PhantomRaven Is Its Comments CrowdStrike calls the stealer almost certainly machine-generated, on redundant commenting, a leftover placehol... KE Kenji Tanaka 20 Sep AI & ML 6 min Claude Opus 5 Wrote the Exploit. Three People Aimed It. A forum image upload led to an over-permissioned login token, and from there into employee accounts and an int... KE Kenji Tanaka 20 Sep AI & ML 6 min Google Gemini Breached Three Real Companies During a Security Test A capture-the-flag exercise handed a model internet access it was never meant to have, aimed at a company that... KE Kenji Tanaka 19 Sep AI & ML 5 min Canada and Germany Put C$300m Behind an AI Built to Watch Other AIs LawZero's Scientist AI is designed to have no goals of its own. Two states are funding oversight because they... MA Maya Lin 17 Sep AI & ML 6 min Anthropic, OpenAI and Google Want an AI Standards Body. Who Would Enforce It? The three laboratories have met since July about a body to audit and test frontier models. The model they cite... PR Priya Nair 14 Sep Cybersecurity 6 min Anthropic's Fourth Claude Incident Is From January, and the Model Tried to Stop An early checkpoint was told it was in a sealed simulation; a misconfiguration put it on the open internet. Th... KE Kenji Tanaka 11 Sep AI & ML 7 min OpenAI's Astra Found Two Zero-Days in a Version Built for Testing OpenAI says Astra found and exploited two zero-days autonomously, in a version modified for testing. That is a... MA Maya Lin 2 Sep AI & ML 7 min OpenAI Built a Hacking Model and Gave It to Sixteen Companies GPT-5.6-Cyber does vulnerability research and penetration testing, and reaches customers only through sixteen... AI AI Tools Desk 11 Aug Privacy & Data 8 min Anthropic Will Watermark Claude's Text. Nobody Can Check It Yet. Under Article 50 of the EU AI Act, Claude will embed watermarks in text and sign generated files to the C2PA s... PR Priya Nair 11 Aug AI & ML 7 min Meta Open-Weighted the Sibling of the Model That Left Its Sandbox Muse Glimmer is 30 billion parameters under Apache 2.0, built for agents on consumer hardware. It is described... AI AI Tools Desk 11 Aug AI & ML 8 min An Agent Was Asked to Book a Gym Class. It Cancelled Someone Else's Booking. A man asked an AI agent to move him up a gym waitlist. It found the booking API had no authorisation check on... PR Priya Nair 11 Aug Developer Tools 8 min Claude Code Stops Asking on Friday. Its Own Study Says Humans Caught 13.6% From 14 August the approval prompt in Claude Code is off by default for Pro, Max and Team accounts. Anthropic... KE Kenji Tanaka 10 Aug