AI Safety news and analysis on Web Pulse: 23587 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Coinkite released firmware 5.6.1 and 1.5.1Q for its Coldcard Mk4, Mk5, and Q hardware wallets after a seed-generation flaw allowed attackers to steal roughly $130 million in Bitcoin. The update requires users to add rand…
Anthropic has launched Claude Mythos 5-powered security scanning in public beta for Claude Enterprise customers, enabling teams to run frontier vulnerability scans without direct model access. The scans connect to a GitH…
OpenAI's GPT-4o hallucinated a non-existent asyncpg API, inventing parameters like health_check_interval and health_check_query for create_pool() and a Pool.health_check() method, which caused TypeError and AttributeErro…
A new study finds that watermarked AI text fools most readers in blind tests, with only 53% of participants correctly identifying watermarked LLM output as AI versus 51% for unwatermarked LLM output, while human text was…
The Future of Life Institute's Summer 2026 AI Safety Index graded Anthropic as the top performer with a C+ (2.66/4.0), followed by OpenAI with a C (2.28), while Google DeepMind scored a C (2.01), Meta a D+ (1.67), and xA…
AI adoption in crypto crime rose 40% over the past year, with TRM Labs' 2026 AI-in-Crime Adoption Index placing overall use at an 'emerging' level of 54 out of 100, up from about 28 in 2024. Scams reached a 'mature' leve…
Anthropic has deployed its most powerful model, Claude Mythos 5, in its security scanner Claude Security to scan codebases for vulnerabilities, provide severity ratings with CWE classifications, and suggest patches. The …
A developer who writes commercial copy for a living has shipped Cencurity, an open-source security gateway that proxies traffic between AI coding agents and LLM providers to mask sensitive data and inspect generated code…
Veracode's Spring GenAI Code Security Update found that 45% of code generated by 150 large language models across 80 tasks contained known vulnerabilities, with security pass rates stuck at about 55% for two years. Weakn…
OpenAI is developing a locked chats feature for ChatGPT that would let users hide conversations behind a PIN or fingerprint, keeping them out of the sidebar, history, and search results. The feature, spotted in ChatGPT A…
A 2022 survey of AI researchers with a 17% response rate found that the majority believed there is a 10 percent or greater chance that human inability to control AI will cause an existential catastrophe. In 2023, hundred…
OpenAI announced Zero Data Retention (ZDR) for its frontier API models on August 19, promising that prompts and responses are deleted after processing with no employee review or training use. The company also previewed P…
A developer has released Grantor, an MCP server that gives AI sub-agents cryptographically bounded permissions instead of full access to the parent's credentials. The tool, available via npx, allows granting limited tool…
A developer building PlannerCritic, an open-source engine where one LLM writes a plan and a second reviews it, found that an adversarial critic prompt caused the reviewer to block plans for completeness concerns rather t…
Aikido Security's survey of 200 CISOs and 200 engineering leaders found that 76% of teams deploy significant changes at least weekly, yet only 21% validate security on every release, highlighting the need for AI pentesti…
A developer analyzed 57 sessions with Claude and found that rules enforced by machine checks had zero violations in 3,131 opportunities, while rules held only in memory were violated 30–45% of the time, with one rule's c…
A developer testing Qwen3.5 4B via Ollama found that enabling chain-of-thought can return an empty response while the thinking field holds content, and that self-critique can degrade accuracy, citing research by Huang et…
A developer highlights the need for pre-execution guards in AI agent pipelines, citing that 72% of organizations now use AI in at least one business function. The developer introduces Agent-Guard, an open-source middlewa…
Start-up noRecognition, founded by cybersecurity researcher Bill Swearingen, has proposed using its anti-AI computer-generated patterns as vehicle wraps to hide cars from Flock's AI-powered automatic license plate reader…
A new method called COPA reduces attack success rates by 6.3× versus the best static baseline and 4.4× on average across lifelong attack streams, while retaining 92% defense on month-old attacks compared to 61% for repla…