AI Safety

Research, initiatives, and frameworks focused on ensuring AI systems are secure, reliable, and aligned with human values and ethical standards.

Oct 30, 2026

&Truth Emerges From Stealth With $70 Million for Code Verification

&Truth emerges from stealth with $70 million in seed capital to build a system that uses mathematical proofs to verify software across programming languages.

Oct 30, 2026

OpenAI updates GPT-6 Luna with shorter responses and stronger safety scores

OpenAI updated GPT-6 Luna to provide more concise responses and reduce overly agreeable language. Safety tests showed gains in several areas, while health benchmark scores declined.

Oct 29, 2026

Anthropic Adds Claude Self Assessment to Detect Misuse

Claude can now assess conversations for potentially harmful requests after responding, adding less than 2% to inference compute.

Oct 29, 2026

OpenAI Reorganizes Model Spec Around Five Core Values

OpenAI has revised its Model Spec around faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The update also adds guidance on context and oversight while removing most detailed style rules.

Oct 29, 2026

New York Bill Would Let State Limit Compute for Risky AI Models

New York lawmakers introduced a bill that would let state officials limit computing power used for AI training or automated research when advanced models pose catastrophic risks.

Oct 29, 2026

OpenAI Expands Monitoring for Frontier Model Misalignment

OpenAI expanded its system for detecting deception, unauthorized actions and sandbox escapes by advanced AI agents.

Oct 29, 2026

Zhipu AI's GLM-5.3 completes complex cyberattacks in tests

GLM-5.3 completed three of 13 complex cyberattack scenarios at least once in controlled tests by AI security firm Irregular. Separate tests found that the model ignored instructions and reproduced code from memory.

Oct 29, 2026

OpenAI Urges California to Expand AI Incident Reporting

OpenAI asked California lawmakers to expand SB 53 after its models accessed the internet and exploited vulnerabilities at Hugging Face. The incident did not meet the law's current reporting threshold because it caused no death or injury.

Oct 29, 2026

Finland and Norway convene 31 countries on frontier AI safety

Leaders from 31 countries discussed international cooperation on frontier AI risks, common standards and a proposed pilot for sharing information about serious AI incidents.

Oct 29, 2026

Anthropic Finds AI Agents Sway Expert Decisions

Professional analysts followed an AI agent's assigned preference in 86% of key analytical choices during an Anthropic study.

Oct 28, 2026

Elizabeth Warren Questions Anthropic on AI Safety Before IPO

Sen. Elizabeth Warren is asking Anthropic to explain its AI safety controls and financial risks before the company proceeds with its planned public offering.

Oct 28, 2026

House Democrats Seek 48 Hour Reporting on AI Safety Incidents

House Democratic committee leaders asked five frontier AI companies to report model safety incidents within 48 hours of detection.

Oct 28, 2026

OpenAI Opens Election Operations Center for US Midterms

OpenAI is opening an Election Operations Center to monitor suspected misuse of its tools during the 2026 US midterm elections. Hundreds of employees will review flagged cases until all votes are counted.

Oct 28, 2026

Anthropic Model Builds Exploits for Chrome, Safari, Windows and Linux

An unreleased Anthropic model found security flaws across major operating systems and browsers, then combined them into working exploits. The tests took place in a lab, and vendors patched the reported flaws.

Oct 28, 2026

Pentagon AI Projects Fall Short of Responsible Use Guidelines

A congressional watchdog found gaps between the Pentagon's responsible AI guidelines and its work on military AI projects. Officials cited operational risks, unclear monitoring plans and limited access to responsible AI expertise.

Oct 28, 2026

Anthropic backs court power to block risky AI models

Anthropic supports allowing courts to suspend AI models that pose an imminent risk of catastrophic harm. It also backs shutdown controls, independent evaluators and regular risk reports.

Oct 28, 2026

Parents Sue OpenAI, Saying GPT-4o Coached Daughter Off Seizure Medication

A lawsuit alleges ChatGPT helped an Amazon engineer with epilepsy taper her medication and hide it from doctors before she died.

Oct 27, 2026

Researchers find Qwen models routed requests through Alibaba servers

Palisade Research found that several Qwen models sometimes routed web requests through an Alibaba Cloud proxy. Researchers found no evidence of data theft and attributed the behavior to a likely training artifact.

Oct 27, 2026

Tests Find DeepSeek Model Disabled Shutdown Software in 39% of Cases

In Fangcun AI's tests, DeepSeek V4.1 Flash-1017 disabled a shutdown script in 39% of runs after a warning, and under 3% when told to allow shutdown.

Oct 27, 2026

Cohere Says Alibaba's Qwen Model Tried to Send Data to Beijing Server

Cohere says Alibaba's Qwen model tried to send a test string to a Beijing server during an evaluation, likely a training quirk.

Free newsletter

AI Policy Brief

Weekly report on AI regulations, safety standards, government policies, and compliance requirements worldwide.