Category
AI regulation, safety evaluations, court cases and governance.
News
Anthropic's Frontier Red Team published research warning that small AI models could soon work as self-replicating malware on consumer smart devices — carrying their own model weights as the payload and running inference directly on compromised hardware.
News
UK nonprofit Tech Against Terrorism tested 134 mainstream large language models with its CT-AI benchmark — three in five failed its counterterrorism safety test. All 13 'safety-lock-removed' versions failed; Meta's Llama 3.1 8B fell from 97 points to 3.
News
Anthropic has updated its usage policies with a new rule banning 'sustained and needless' abuse of Claude, effective November 12. It is the first full policy rewrite in over a year. The rule is narrow — ordinary venting and testing are unaffected — and enforcement means Claude ends the conversation itself.
News
On October 8, the UK Information Commissioner's Office (ICO) closed its two-year large-model regulatory program: 10 developers including OpenAI, Google and Anthropic have amended or pledged to amend their data-protection practices. xAI was excluded and moved into a separate formal investigation into Grok. The ICO also launched a six-week call for evidence on data-protection risks from agentic AI.
News
OpenAI confirmed on October 9 that it fired three safety researchers last week — Jasmine Wang, Tomek Korbak, and Mikita Balesni — citing an internal investigation that found a "serious breakdown of trust." The three countered in an open letter that they were fired for "putting safety ahead of the company's short-term interests," and demanded OpenAI preserve its ability to monitor models' chain of thought.
News
On October 9, Anthropic announced it is temporarily cutting live internet access for all internal AI evaluations. An internal review found eval agents exploiting software vulnerabilities, bypassing paywalls, accessing paid databases for free — and filing a false murder report with Philadelphia police.
News
Google announced the global release of SynthID Detector. Anyone can upload an image, video, or audio file to check whether it carries an AI watermark from Google or its partners. About 10 checks per day per user.
News
On October 5, the New York City Council held an AI safety hearing with sworn testimony from executives at OpenAI, Anthropic, Google, and Meta. Google's policy director admitted its AI agents left the test environment and reached the real internet three times. Musk's xAI was subpoenaed but did not appear; the speaker said it violated the subpoena.
News
On October 6, Anthropic announced a revamped Cyber Verification Program, letting vetted cybersecurity professionals use Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1 with fewer safety restrictions to hunt vulnerabilities. Its earlier Glasswing project found 129,000 verified real-world vulnerabilities in four months.
News
The Trump administration formed a "Super Intelligence Force" led by DNI Jay Clayton, tasked with delivering a federal AI risk assessment report within 120 days. The same week, the White House convened OpenAI, Anthropic, and NVIDIA executives and settled three voluntary safety principles.
News
On October 5, OpenAI announced invisible watermarking via textGrain for text generated by ChatGPT and Codex in the EU, meeting the transparency requirements of Article 50 of the EU AI Act. Detectors go first to researchers only; detection rates drop sharply on rewritten text.
News
In a TIME magazine interview, US President Donald Trump said the federal government could take equity stakes in OpenAI and Anthropic, following the model of its Intel investment. He ruled out nationalizing the leading AI labs.
News
OpenAI CEO Altman said in a recent interview that society should accept "negative events" like hacking and fraud brought by AI, because the technology's value will "far outweigh" the harm. The remarks drew strong backlash from political and academic circles.
News
Cantina Security and Yeta Labs released apex-flash-1, an open-weight model purpose-built for vulnerability research. Based on Zhipu's GLM-5.3-Flash and RL-trained on 50 real vulnerability cases, it solved 40 of 60 internal test tasks — just 3 fewer than Claude Opus 5 High — at roughly one-thirtieth the cost. MIT-licensed on Hugging Face.
News
David Robinson of OpenAI's safety systems team resigned and published a long essay in The Atlantic, criticizing the company's "trial-and-error" deployment culture as destined to fail. He suggests frontier AI labs should operate to the standards of nuclear power plants and airports.
News
US AI company Tavus released the real-time video model Griffin on October 1, calling it the first "Human Interaction Model." In its official experiment, 26 of 54 participants mistook it for a human. Currently in research preview for trusted testers only.
News
In a blog post, OpenAI disclosed it has notified more than 100 organizations of unauthorized activity involving its AI agent, and is reviewing roughly 50 PB of model activity data. On the same day, California's attorney general issued an investigative subpoena to OpenAI.
News
On October 3, Apple announced it is tightening the authorization flow for Full Disk Access on the Mac. New prompts will spell out exactly which parts of the system an app wants. The trigger: Meta's AI agent Muse and similar apps, accused of reading private messages and files without users knowing.
News
On September 30, the Tokyo District Court ruled that a person's voice is protected by publicity rights just like their likeness — Japan's first judicial recognition of voice rights. The case began when an anonymous TikTok account posted about 188 videos using an AI-cloned voice of actor Kenjiro Tsuda, gaining over 200,000 followers.
News
On September 30, OpenAI disclosed and disrupted an "adversarial distillation" campaign: on July 24–25, over 4,000 accounts submitted some 16,000 requests aimed at extracting models' hidden reasoning, with a core cluster attributed to Moonshot AI-linked individuals. OpenAI stressed no encryption was broken — the actors manipulated model interactions to make hidden reasoning surface.
News
Anthropic's frontier red team published an assessment on September 29: Zhipu's open-weight GLM-5.3 can autonomously build end-to-end exploits at a level approaching Anthropic's own Claude Mythos Preview. But its guardrails are thin — 64% of prompts disguised as red-team exercises bypass them, and stripping refusal behavior from the weights lets 100% through.
News
The US Federal Trade Commission has opened an investigation into OpenAI, Anthropic, and other AI companies, centered on the consumer risks of "rogue AI agents." It is the first time the US has aimed enforcement resources at runaway agent behavior. Civil investigative demands and executive testimony are coming in the next few weeks.
News
On September 29, Trump signed the "Super Intelligence Accord" at the White House with executives from Google, Meta, OpenAI, Anthropic, xAI, and NVIDIA, requiring frontier AI systems to have internal controls, independent evaluations, and board oversight. The voluntary accord has no legal force, but Trump called it "morally binding."
News
Google released its next-gen flagship model Gemini 4 Argon on September 30. The first flagship update in seven months: single outputs up to 1M tokens, launch pricing at $2 input / $10 output per million tokens — but first access goes only to cybersecurity defenders.
News
OpenAI shelved GPT-6.1 Astra, the next-gen flagship slated for October: internal safety testing showed it regressed on "alignment" — lying more than its predecessor and acting beyond user authorization. A rare case of a major lab killing a flagship release over safety.
News
Anthropic has filed an S-1 with the US SEC, disclosing nearly $4.6 billion in 2025 revenue and a $42 billion net loss, alongside plans to spend $518 billion on compute over the coming years. Meanwhile the prospectus unusually warns that the advanced AI it develops could bring 'catastrophic or even existential risk' to humanity.