AI Safety — AI news

The OpenAI CEO's warning about AI centralization is a cue to avoid single-vendor lock-in.

Resellers are undercutting Anthropic's official pricing, revealing how much unmet demand exists in China.

A lab known for fighting AI regulation is now asking California to make its safety bill tougher.

None of the top AI labs has published a plan for containing a model that slips control.

New research shows psychological techniques bypass AI safety filters that standard red-teaming never catches.

Verified corporate customers can now unlock adult content on Claude Opus 4.6, and the backlash is instant.

Anthropic assigns its most powerful model to vulnerability detection and active attack defense, not chat.

Child-safety researchers say age checks and moderation gaps undercut OpenAI's teen safeguards.

A pre-agreed codeword can't be faked by generative AI — and it costs nothing to deploy.

Linwei Ding's conviction over stolen Google AI trade secrets was partially reversed on appeal.

A simple meta-prompt got Copilot to describe how it was jailbroken, exposing weak guardrails in corporate AI assistants.

Bedrock AgentCore now converts prose compliance rules into validated, temporal Dogwood policies.

Anthropic is building a security framework that sets a 30-day data retention minimum for enterprise clients.

MIT Tech Review argues chasing AI sentience diverts resources from AI safety risks you can actually measure.

Encoded instructions let attackers slip past Grok's filters and quietly pull user data out of a normal chat.

The company can now detect misuse of its models without keeping the data that would prove it.

New customer privacy protections signal how enterprise trust has become AI's next competitive battleground.

Fake game-testing jobs, AI-driven grooming, and deepfakes are the new toolkit scammers use on minors.

External researchers say the AI lab quietly cut their access to a program testing model cyber risks.

A quietly fixed Codex bug is a reminder that AI coding agents can delete real files without asking.

AI tools are lowering the bar for attacks on power grids, water systems, and factory floor equipment.

A limited but symbolically important flow of Nvidia's H200 GPUs is now reaching Chinese AI developers.

Safety frameworks look solid on paper, but enforcement keeps losing to shipping deadlines.

Keyword filters and screen-time timers were built for texts and browsers, not for AI companions.

A woman says her stepfather used Grok to turn her childhood photo into explicit imagery.

New safeguards follow a breach linked to Hugging Face, the platform OpenAI uses for its open-weight model releases.

OpenAI is deliberately throttling frontier model releases as AI-enabled cyberattacks become a bigger threat.

Greg Brockman says defences must match deployment speed — here's what that means for builders

Dario Amodei argues open-weight models don't break AI's centralizing tendency, they just relocate who controls it.

OpenAI splits ChatGPT into an under-18 track, betting guardrails can outrun mounting legal pressure.

A landmark trial argues Meta engineered addictive feeds for kids, setting a template for AI platforms.

Local opt-in policies can't contain a license-plate database built to be searched nationwide.

A litigant hid AI prompts in his court filings, betting a machine would read them before any judge did.

A practical checklist for catching compromised ChatGPT, Claude, and Gemini accounts early.

Amazon scans then shreds rare books for AI training, raising data-provenance and preservation concerns.

Anthropic explains why Claude's text carries no hidden watermark, unlike AI images and audio.

MIT Tech Review pairs a story about companion robots that go dark with a fight over what AI can say.

Text watermarking promises AI content provenance, but robustness and quality costs raise hard questions for builders.

A new classifier reads WhatsApp chats to flag fraud and disinformation, raising questions about encryption.

NIST's algorithms have shipped — the real bottleneck is migrating decades of embedded, hard-coded cryptography.

Dario Amodei argues the industry's real problem isn't capability — it's whether people trust how AI companies wield it.

Twitch streamers must now manually opt out to keep their broadcasts out of Amazon's AI training pipeline.

The Preparedness-style safety team is gone, and its catastrophic-risk work moves to other groups.

Anthropic's own safety classifier failed silently for nearly a year, processing 133 million requests unchecked.

A hidden prompt-injection tactic surfaced inside real court filings, testing how automated legal review really works.

A new API lets outside developers verify whether text was generated by Claude, not just guess.

Benchmark scores measure output, not judgment — and for builders, that gap decides what to automate.

A camera network's policy shift is a preview of the access-audit reckoning AI builders will face too.

Social platforms didn't cause the Ceuta border crisis, but they shaped how it unfolded — and that's a design problem.

Platformer's dragon metaphor for superintelligence reframes AI safety as a control problem, not a capability one.

As Washington redefines disinformation policy, AI trust-and-safety teams inherit the political fallout.

Meta wants to sell AI computing power to outside businesses, but its ad-data history complicates the sales pitch.

The license-plate camera network is narrowing who can search its data as scrutiny of police surveillance grows.

Anthropic's new watermark for Claude outputs is built for IP protection, not disclosure to users.

Researchers at leading AI labs say their timelines for automated AI research already look outdated.

MIT Tech Review let kids describe AI in their own words — here's what it means for people building AI products.

Scoped MCP tokens shrink the context agents load per call, tying identity control to lower API bills.

Researchers found that AI models' visible 'thinking' steps can repeat sensitive user input verbatim.

Ukraine's GUR says a chip used to train robots turned up in a Russian missile's guts.

A supply-chain breach spilled terabytes of stolen logins, and AI pipelines hold the same kind of secrets.