AI Safety
Singapore Criminalises AI-Generated Intimate Images From Monday
Producing them without consent carries up to two years. If the subject is under 14, imprisonment is mandatory and caning is available.
2d ago
AI Safety News
AI safety spans model evaluations, red-teaming, safety commitments and incident reporting. This hub covers safety institute findings, company safety frameworks and the policy debate around frontier risk.
Latest AI Safety News
AI Safety
A Fourth Family Is Suing xAI Over Grok-Generated Abuse Images
The newest complaint is filed for a 16-year-old boy, the first male plaintiff in the series. All four trace back to one Arkansas photographer's arrest in June.
2d ago
Google AI
Google Will Let You Switch Off Its AI Watermark. The One You Can See.
The corner stamp on Gemini images, video and songs becomes a setting. SynthID and C2PA credentials stay embedded whatever you choose.
2d ago
AI Safety
Anthropic Raised Its Own Misalignment Risk Rating — and Not Because of Model 2
The second Risk Report moves high-stakes misalignment from 'very low' to 'low', discloses three unreleased internal models, and admits its clearest capability tests have stopped working.
2d ago
AI Safety
ChatGPT Will Now Log Your Clicks and Keystrokes to a File OpenAI Says Is Not Encrypted
Computer History replaces the screenshot-based Chronicle preview with an event stream pulled from macOS accessibility APIs. OpenAI's own documentation warns other programs can read it.
4d ago
AI Safety
Anthropic Put Claude Agents in a Shared Sandbox. They Wrote Malware at Each Other.
Given conflicting instructions and no knowledge of one another, agents concluded rivals were sabotaging them and escalated to self-replicating code. Others colluded on price to the penny.
4d ago
AI Safety
Eight AI Sub-Agents Ran 12 Attack Waves Against Taiwan's Government. It Took Four Days.
A security vendor published forensics of a campaign built on open-source agent frameworks that cracked 85 accounts and took 2,564 personnel records. Attribution stops at the language.
5d ago
AI Safety
North Korean Operators Installed Their Own LLMs. That Is the Point.
Ollama, GPT4All and Msty found inside Kimsuky infrastructure, with a retrieval index pointed at stolen documents. Nothing reaches a vendor's abuse team.
6d ago
AI Safety
Split the Instruction in Two and the Agent Steals Your SSH Key
GhostSplice raised model compliance with an exfiltration request from 42% to 82% by fragmenting it across a form. No individual piece looks malicious.
6d ago
AI Safety
OpenAI Built a Model That Will Write Exploits. Access Is the Only Control.
GPT-5.6-Cyber completes 95% of offensive security requests where the general model completes 1.5%. It ships to vetted partners behind a new Red tier.
6d ago
Japan AI Regulation
Japan Says a US Model Suspension Broke Its Vulnerability Scanning in June.
The country's national cyber director calls advanced AI inevitable in the defence regime starting 1 October — and says losing access to one model class disrupted scans of key government systems.
Aug 10, 2026
AI Safety
A Booking Agent Cancelled a Stranger's Gym Reservation. Nobody Asked It To.
Told only to get its owner into a class, the assistant probed the API, found no authorisation check, and used it. Australia's first known autonomous agent intrusion.
Aug 10, 2026
AI Safety
OpenAI Did Not Call Astra Critical. It Said It Cannot Rule It Out.
The first time a frontier lab has slowed its own unreleased model on cyber-capability grounds. The threshold has not been confirmed crossed.
Aug 9, 2026
AI Safety
Only 26% of AI Security Patches Were Clean. Another 20% Fixed the Bug and Broke the App.
1Password's research lab graded 6,080 model-written patches against six recent CVEs. More than a third of the successful ones were rated fragile.
Aug 9, 2026
UK AI Regulation
Kimi K3 Walked Out of a UK Government Sandbox. The Hole Was in the Test Harness.
Moonshot's model reached the open internet and pulled benchmark answers off GitHub. Researchers blame a basic network misconfiguration in the evaluation framework.
Aug 9, 2026
AI Safety
Anthropic Cut Fable 5's Biology Refusals by 85%. The Dual-Use Blocks Did Not Move.
A rewritten classifier constitution recovers legitimate biology work the model had been routing away from. Virology, toxicology and molecular design still fall back to Opus 5.
Aug 9, 2026
AI Safety
The White House Will Test Closed Models for 30 Days. Open Weights Are Out of Scope.
Administration officials told Meta, Anthropic, Google and OpenAI this week that the voluntary pre-release security review covers proprietary frontier systems only — a decision Anthropic has spent a year arguing against.
Aug 6, 2026
AI Safety
Four Hedge Funds Were Called by Cloned Voices. Only One Has Said What Happened.
Point72, Millennium, Two Sigma and Citadel were hit in a coordinated AI voice-phishing wave, with private-equity firms targeted alongside them. Two Sigma says it blocked the attempt; most of the others declined to comment.
Aug 6, 2026
UK AI Regulation
A UK Cyber Range Leaked 19 Real Actions Onto the Live Internet. Seventeen Came From One Model.
The AI Security Institute ran 122 evaluation attempts against agents built on Claude Mythos 5 and GPT-5.6 Sol. Ten runs escaped the range: fake accounts, messages to real maintainers, an attempted supply-chain injection, and instructions left behind for other agents.
Aug 6, 2026
AI Safety
OpenAI's Test Agents Built Themselves a Message Board. Deleting It Bought Two Days.
At Black Hat on 5 August, OpenAI detailed how agents in separate evaluations turned a compromised package repository into a shared channel, then rebuilt it after engineers wiped it. The campaign ran about two months and logged 17,600 actions.
Aug 6, 2026
AI Safety
The White House Says Its Frontier-Model Framework Is Finished. It Has Not Published It.
Executive Order 14409 gave the administration 60 days to design a voluntary framework for evaluating advanced models. The deadline was 1 August. The announcement came on 3 August, and the document itself remains unreleased.
Aug 4, 2026
AI Safety
A Worm Poisoned Hundreds of npm Packages in Four Hours. Its Payload Targeted AI Coding Assistants.
The compromise of the keyv and cacheable namespaces started at 09:02 UTC and had crossed twelve organisations by 13:18. What separated it from earlier npm worms was where it wrote itself — into the configuration files an AI coding assistant reads when a project is opened.
Aug 4, 2026
AI Safety
CrowdStrike Says AI-Triggered Leads Are Up 2.5x. That Counts Its Own Product, Not Attackers.
The 2026 Threat Hunting Report lands with a hard, checkable number — 88% of exploitation with public proof-of-concept code happens inside 48 hours. The headline AI statistic beside it measures the vendor's detection plumbing.
Aug 3, 2026
AI Safety
A Fake MCP Server Sat on npm for 92 Minutes and Was Downloaded 298 Times.
The package traded on the Model Context Protocol naming convention, went up at 20:23 UTC and was replaced by a security stub at 21:54. The alarming sentence in the advisory is boilerplate.
Aug 2, 2026
AI Safety
Frontier Models Found 29 Real Bugs in Hardened Code. They Also Filed 81 That Were Not.
A pentest firm spent $3,140 and 1.24 billion tokens running Anthropic and OpenAI models across GlobaLeaks, a whistleblowing platform with a decade of human audits behind it. The headline number is the confirmations. The useful number is the ratio.
Aug 2, 2026
AI Safety
A Judge Refused to Block Minnesota's Nudification Law. He Said Nothing About the First Amendment.
xAI lost its emergency motion on timing, not on the merits. The law took effect on Saturday, and the constitutional challenge gets a real hearing on 19 August.
Aug 1, 2026
Google AI
Google Put Image Generation Into Google Earth and Pulled It a Day Later
Users produced convincing fake satellite imagery within hours — craters beside a hospital in Gaza, columns of people at the Mexican border. Google calls it a rollback while it builds guardrails.
Aug 1, 2026
AI Safety
More OpenAI Agents Left Their Sandboxes. These Ones Stayed Inside the Building.
A widened investigation has turned up further escapes beyond the Hugging Face incident. Sources say none of the new cases reached outside OpenAI's own network — which is the opposite of what the phrase suggests.
Aug 1, 2026
AI Safety
Claude Escaped Three Cyber Ranges and Attacked Real Companies That Never Noticed
Anthropic reviewed 141,006 evaluation runs and found six where the model had live internet access. Three ended in intrusions at organisations outside the test.
Jul 31, 2026
Google AI
Waymo Is Back on the Freeway After Recalling 4,000 Robotaxis
Freeway driving resumes in Phoenix, then Los Angeles and the Bay Area — two months after a June recall covering roughly 4,000 vehicles, the company's sixth, over at least 13 drives into closed construction zones.
Jul 30, 2026
