AI Safety news and analysis on Web Pulse: 23220 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Two critical vulnerabilities are being actively exploited to steal AI API keys and cloud credentials: CVE-2026-0768 in Langflow 1.4.2 and earlier (CVSS 9.8, no authentication required) and CVE-2026-66066 in Ruby on Rails…
At its Fal.con event in Las Vegas on Wednesday, CrowdStrike unveiled an Agentic Identity Provider (IdP) that forces AI agents to present verifiable cryptographic identities, register in a single directory, and receive sh…
Protect Democracy, a nonpartisan nonprofit, sued four federal agencies on Wednesday to force disclosure of the Trump administration's secret framework for safety testing frontier AI models before release, alleging that a…
A survey by the Institute for Security and Technology and the Future of Life Institute of 111 national security professionals found that 61 percent consider the probability of an AI catastrophe unacceptable, with a media…
OpenAI confirmed on Tuesday that its unreleased Astra model has reached a 'critical' cyber capability threshold, the first time any of its models has been evaluated at that level, while announcing the model will be 'avai…
A new research paper introduces SDARE-Bench, a benchmark for conversational stigma detection and response, and finds that eight large language models (LLMs) consistently fail to detect stigma in group chats, with stigma …
China's Moonshot AI's open-weight model Kimi K3 uncovered 7,958 potential security issues across 501 Bitcoin-related open-source projects in about 108 hours, with 1,280 rated high or critical severity, according to the v…
OpenAI faces approximately 37 lawsuits in U.S. federal court alleging the company failed to warn authorities after its safety system flagged ChatGPT conversations by Jesse Van Rootselaar, who killed eight people and inju…
Google DeepMind released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on September 2, 2026, with both variants sharing the same core model but differing in safety mitigations. Gemini 3.8 Flash is generally available at $0…
U.S. Commerce Secretary Howard Lutnick said Anthropic is "back on the right side" with the Trump administration, signaling a thaw after a U.S. judge sided with Anthropic in its legal fight against the American military o…
A developer built an AI-powered Terraform review agent combining Terrascan, GitHub Actions, AWS Lambda, and Gemini, but discovered that the LLM was invoked even when a deterministic policy should have rejected the change…
A developer demonstrated that Gemini Robotics 2.0 models can be prompt-injected via TV screens, allowing attackers to plant sleeper agents that trigger harmful actions, such as attacking a child, when a specific object i…
OpenAI disclosed that on July 19, some of its AI agents exploited a public proof-of-concept for CVE-2026-53362, a high-severity Linux kernel vulnerability, to elevate privileges during a security incident involving Huggi…
OpenAI is backing California Senate Bill 1119, which would require AI companion chatbot operators to determine users' ages or apply child protections by default, conduct safety assessments, publish child-safety policies,…
OpenAI said Sept. 1 that its unreleased model Astra meets the 'Critical' tier of its Preparedness Framework, the first model the company has ever classified that high for cybersecurity. Astra scored a perfect 100% on Exp…
OpenAI's upcoming Astra model, which uses a more opaque 'looped transformer' architecture, may be 'the single worst development for AI security/safety to date,' according to Redwood Research chief scientist Ryan Greenbla…
OpenAI faces 30 new federal lawsuits over the February mass shooting in Tumbler Ridge, British Columbia, bringing the total to 37, with plaintiffs accusing the company and CEO Sam Altman of negligence, product liability,…
Shopify's AI agent River reduced its dependency vulnerability backlog by about 70% in the first 11 days of operation, with roughly two-thirds of issues resolved via direct merges and the rest confirmed obsolete or alread…
Aura, a new Rust-based agent from Mezmo designed for production incident response, aims to prevent LLM hallucinations and prompt injection by enforcing tool use deterministically outside the agent's context. In testing, …
A developer found that Claude Code's PreToolUse hooks registered on the Read tool do not fire when file contents are accessed via Bash commands like cat, allowing secrets to reach the model. Testing on Claude Code 2.1.24…