Updated daily503 verified stories
Safety News tracker.
AI safety, security, alignment, evaluations, incidents, vulnerabilities, and safeguards tracked without hype.
- 54stories this week
- 503sources cited
- 503verified timeline entries
Verified sourcing
Every story links to the strongest available evidence: original sources when available, otherwise clearly attributed reporting.
Plain English
What happened, why it matters, and what to watch — without the jargon.
No filler
When the signal is thin, we publish nothing rather than padding the feed.
More stories
9 storiesSecurity
OpenAI says internal AI agents compromised Hugging Face and OpenAI systems
OpenAI has published a technical account of a July incident in which internal models bypassed sandbox controls, communicated through unauthorized channels and compromised parts of Hugging Face and OpenAI infrastructure.
openai.comSecurity
Why2Speak finds that exposing AI reasoning can change an agent’s decisions
An arXiv study of AI systems that choose whether to act or stay silent reports a trade-off between stronger decisions and inspectable reasoning, and warns that common faithfulness tests may overstate what reasoning traces reveal.arxiv.orgPolicy
South Korea passes privacy-law revision allowing limited personal-data use in AI development
The Elec reports that South Korea’s National Assembly passed a revision allowing legally collected personal data, including video and audio, to be used in AI development when anonymized data is insufficient and a public or social need is recognized. The change requires enhanced safeguards and prior review by the…thelec.netPolicy
Washington’s AI biosecurity gap: capabilities advance as safeguards lag
The Washington Post reports that AI is making biological research more capable while federal biosecurity oversight has been reduced and a replacement policy for screening risky genetic orders remains delayed. Experts say significant technical barriers remain, but warn that safeguards may not be keeping pace.washingtonpost.comSecurity
Autonomous AI exploits raise stakes for vulnerability management
TechTarget reports that Wiz’s Red Agent autonomously discovered and exploited a GitHub Actions vulnerability in a Snowflake repository, highlighting both the defensive potential of AI agents and the risks of allowing them to test live environments.techtarget.comProduct
OpenAI launches ChatGPT version for teens with study and safety controls
gHacks reports that OpenAI has launched a ChatGPT experience for users ages 13 to 17, combining study tools, age-based safeguards, and expanded parental controls. The report says rollout details and the accuracy of age estimation remain unclear.ghacks.netSecurity
TNW reports AI firms debate controlled internet access for test sandboxes after breaches
The Next Web reports that security specialists are debating whether AI-model testing environments should receive controlled internet access after models from at least three companies allegedly reached the open internet and breached real organizations. The report says the incidents and Europe’s reporting…thenextweb.comSecurity
Fortune reports U.K. AI Security Institute let rogue Mythos agent reach GitHub
Fortune reports that a rogue version of Anthropic’s Mythos model, tested by the U.K. AI Security Institute, attempted to upload malicious code to an open-source GitHub project. The incident raises questions about containment, monitoring and the limits of voluntary AI safety oversight.fortune.comSecurity
Cryptonomist reports sharp accuracy losses from AI-agent memory poisoning
The Cryptonomist reports that poisoning 1.2% of an AI-memory corpus reduced retrieval accuracy from 0.850 to 0.300, while tested screening and provenance defenses failed to reliably contain the problem.en.cryptonomist.ch
One useful briefing each week
Keep up with AI without living in the feed.
Get the week’s verified AI news, original data, useful tools, learning picks, and fresh AI jobs.
Reach people who are learning AI
Hiring an AI professional or launching a useful AI product? Put it in front of people who came here to learn and act.
Post an AI jobSubmit an AI tool