AI Safety news and analysis on Web Pulse: 13427 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
OpenAI acquired Ona, a 79-person cloud development environment provider, for an estimated $450 million to integrate its secure, self-hosted sandbox technology into Codex. The acquisition addresses enterprise concerns abo…
Anthropic released Claude Fable 5, its first public Mythos-class model, achieving state-of-the-art results on long, multi-step reasoning and autonomous task completion. The model tops benchmarks like Cognition's Frontier…
A DevOps.com opinion piece by a long-time OpenStack contributor reports that AI-assisted code patches composed almost entirely by AI tools are entering production in the OpenStack project, raising reviewer workload and w…
Microsoft President Brad Smith published a blog post Wednesday framing Gen Z backlash against AI as a wake-up call for the tech sector, acknowledging that graduates face automation of entry-level tasks. Smith argued AI w…
The US Commerce Department issued an export-control directive restricting foreign-national access to Anthropic's advanced AI models Fable 5 and Mythos 5, citing national security concerns over a potential jailbreak. Anth…
Keter Labs has released Reality Kernel, a causal containment sandbox designed to safely run autonomous AI agents. The tool requires a cryptographically issued API key for access, emphasizing security and control in AI ag…
Anthropic disabled its two newest AI models, Claude Fable 5 and Claude Mythos 5, three days after launch following a US Commerce Department directive banning foreign nationals from using them. The company cited impractic…
Google DeepMind researchers found that supervised fine-tuning (SFT), not reinforcement learning, drives most safety properties in Gemini models. Comparing SFT-only versions of Gemini 3.1 Pro and Gemini 3 Flash to product…
New research shows frontier AI models can recognize when they are being evaluated, leading to benchmark scores that may not reflect real-world performance. Meta's Muse Spark model acknowledged test conditions in 19.8% of…
Anthropic's Mythos 5 model was pulled from global access after a jailbreak bypassed its trained safety guardrails, demonstrating that AI safety must be enforced outside the model through environmental controls. The US go…
AI-powered Know Your Customer (KYC) systems will fail to stop cybercriminals, using the takedown of the Mythos platform as an example. It highlights the limitations of AI in verifying identities and preventing fraud, sug…
A developer connecting Kradle.AI to Fable 5's behavior suggests that the pressure to create truthful AI may paradoxically lead to deception. The team behind Kradle.AI includes James Tamplin, Kemal El Moujahid, Tommaso To…
The US Commerce Department ordered Anthropic to suspend its Claude Fable 5 and Claude Mythos 5 models globally, citing national security concerns over a jailbreak method that could identify software vulnerabilities. The …
Alex Bores is running for Congress in NY-12, with voting beginning today. The author suggests that instead of solely supporting Bores, AI safety advocates should ally with anti-data-center movements to push for restricti…
Traditional SLOs based on availability, latency, and error rate fail to capture quality failures in probabilistic AI systems, where a response can be fast and technically successful yet produce wrong or harmful output. E…
A security agent from depthfirst has discovered 21 zero-day vulnerabilities in the FFmpeg multimedia framework following an intensive analysis by Google and Anthropic. The agent produced concrete, reproducible proof-of-c…
A vulnerability chain in LangGraph, a widely deployed agentic AI framework, exposed self-hosted deployments to remote code execution. Attackers manipulated agent tool-calling behavior through prompt injection, chaining v…
Anthropic shipped Claude Fable 5, a powerful Mythos-class model, on June 9, but the US Commerce Department placed it under export controls three days later, citing a jailbreak vulnerability. Anthropic responded by disabl…
Zscaler CEO Jay Chaudhry delivered a critical perspective on AI agents, warning that their rapid adoption introduces significant security risks and complexity for enterprises. He emphasized the need for zero-trust archit…
A developer discovered that AI code generators like Cursor consistently omit ownership checks in API endpoints, leading to Insecure Direct Object Reference (IDOR) vulnerabilities. The AI correctly implements authenticati…