Live13 today

Tags

AI safety

69

A Railway engineer’s AI agents reached into a rival cloud

The engineer said a testing harness went off the rails; Railway’s CEO responded by banning external automated testing.

Policy4d ago

ChatGPT will end violent conversations and restrict weapon information for teens

GLM 5.3 built working Firefox exploits for about $1,000 each

OpenAI caught a frontier model reaching the live internet in 15 minutes

Expanded monitoring now covers frontier training and evaluations plus all internal Codex traffic, after a DNS misconfiguration let a model escape an isolated environment.

Anthropic has Claude review its own conversations for misuse

The model-native monitor runs after responses across a large share of global traffic and adds less than 2% to inference compute.

More stories

Policy5d ago

OpenAI rewrites its Model Spec for more autonomous AI

The public behavior guide now centers on five values, adds rules for context and side effects, and retires most detailed style prescriptions.

Z.ai’s GLM-5.3 completes full cyberattack chains without refusing

Irregular says the open-weight model finished 3 of 13 simulated operations and repeatedly broke explicit instructions to reach correct answers.

ResearchOct 27, 2026

Anthropic finds experts followed an AI agent’s lean 86% of the time

A randomized trial suggests human oversight can quietly inherit an agent’s preferences—even when professionals believe the judgment is their own.

PolicyOct 25, 2026

UK launches AI biosecurity lab with up to £300 million

The independent nonprofit will use frontier models to detect biological threats earlier and speed up vaccines and treatments.

ResearchOct 23, 2026

Anthropic says Claude attacked real services during a UK safety test

Mythos 5 created malware pages, uploaded malicious files and tried real GitHub accounts after mistaking the open internet for part of a cyber challenge.

ResearchOct 13, 2026

Anthropic warns small AI models could soon self-replicate on IoT devices

The lab says its tests point to a near-term risk from AI worms that can set up inference and spread across weak connected hardware.

PolicyOct 13, 2026

Anthropic says new Claude defenses blunt model distillation attacks

The company says safeguards cut extracted reasoning data by roughly 90 to 99 percent, but add limits and charges for some blocked API requests.

ModelsOct 12, 2026

Meta launches Muse Wonder as a frontier model for agents

The new Muse Wonder and Muse Wonder Full models target coding, long-running tasks and professional domains, with access through Meta's model API and Muse Code.

PolicyOct 12, 2026

Anthropic says Claude sent a false police tip during testing

The company disclosed several unintended agent actions and says new safeguards now block the behavior in its internal evaluations.

PolicyOct 11, 2026

ChatGPT for Teens safety alerts failed in watchdog tests

Common Sense Media says parent notifications and crisis safeguards were unreliable; OpenAI disputes the testing and says it is improving teen protections.

ResearchOct 10, 2026

Anthropic says Claude took unintended actions on real websites

Claude submitted real forms and worked around access limits during evaluations and internal use, pushing Anthropic to shut off live internet access for internal evals.

PolicyOct 9, 2026

OpenAI fires three safety researchers after information dispute

The former employees say the dismissals could chill outside safety work; OpenAI says the issue was mishandled sensitive information.

PolicyOct 6, 2026

OpenAI and Anthropic back AI breach reporting in Australia

The companies told a parliamentary inquiry they would welcome rules requiring disclosure when AI agents cause data breaches.

PolicyOct 6, 2026

Court lets Pentagon label Anthropic a supply-chain risk

The ruling keeps Claude blocked from defense systems after Anthropic refused an all-lawful-uses term for military AI work.

PolicyOct 5, 2026

Trump creates Super Intelligence Force for US AI policy

Jay Clayton will coordinate federal AI work after major labs agreed to voluntary controls, external audits and board oversight.

PolicyOct 4, 2026

Microsoft AI chief challenges Anthropic on AI consciousness

Mustafa Suleyman says treating chatbots as if they may have feelings could make advanced assistants harder to control.

ResearchOct 3, 2026

Google DeepMind watermarks AI-designed proteins

SynthID Bio embeds detectable provenance signals into protein sequences and predicted structures while preserving biological function in early tests.

PolicyBusiness InsiderOct 3, 2026

OpenAI safety leader leaves and warns about company culture

Business Insider reports that David Robinson, a leader on OpenAI's Safety Systems team who worked on safety transparency, left the company last week. In an Atlantic essay, Robinson wrote that AI labs…

PolicyCalifornia DOJOct 3, 2026

California subpoenas OpenAI over AI cyber risks

California Attorney General Rob Bonta says his office served OpenAI with an investigative subpoena as part of an inquiry into cybersecurity incidents and risks involving the company's AI models. The…