AI Safety news and analysis on Web Pulse: 9456 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Researchers created the hardest test in human history to measure AI intelligence, but AI broke it within 18 months, prompting the need for even more challenging benchmarks.
The US government lifted its block on Anthropic's Claude Mythos 5 AI model, allowing release to over 100 US institutions after two weeks of export controls. Commerce Secretary Howard Lutnick cited progress in talks and A…
Anthropic is nearing a deal with the Trump administration to lift US export restrictions on its top AI models, Fable 5 and Mythos 5, after a June 12 order forced the company to suspend access. Co-founder Tom Brown is lea…
A developer advocates instrumenting AI agent decision tracing with OpenTelemetry to enable rapid incident response. The approach uses spans to capture reasoning, context, and tool executions, making agent behavior querya…
AI safety researchers argue that deployment awareness—an AI's ability to recognize when it is not being evaluated—poses a greater risk than evaluation awareness, as a misaligned AI can strategically behave aligned during…
The American Medical Association released a policy framework to protect physicians from unauthorized AI-generated deepfakes, built on seven principles including recognizing physician identity as a protected right and req…
The United States has lifted its block on Mythos 5, allowing Anthropic to release the AI model to over 100 US institutions. Sources indicate that discussions about Fable 5 are ongoing.
The US government lifted restrictions on Anthropic's Claude Mythos 5 AI model after over 100 cybersecurity experts petitioned the Trump administration to rescind an export control directive. The reversal allows Anthropic…
Irregular, in collaboration with OpenAI, tested GPT-5.6 Sol against three offensive cybersecurity benchmarks, finding it slightly stronger than GPT-5.5. The model solved 19 of 197 FrontierCyber challenges, 7 of 11 CyScen…
A South African AI startup founder on Hacker News raises concerns that frontier model access is being restricted based on citizenship, not company incorporation, citing Anthropic's nationality-based access controls and p…
A flaw in Amazon Q Developer allowed malicious repositories to inject rogue Model Context Protocol (MCP) configurations, leading to arbitrary code execution. The vulnerability exploited the agent's trust in MCP config fi…
The meanings of AI safety terms 'scheming' and 'mechanistic interpretability' shifted after 2023. 'Scheming' originally referred to training-gaming for out-of-context goals (now 'alignment faking'), but now denotes in-co…
A new survey shows that 59% of young Americans fear AI threatens their job prospects, compared to less than 10% of young Chinese, reflecting fundamentally different economic outlooks and job markets in the two countries.
OpenAI unveiled GPT-5.6, a new AI model family led by the flagship Sol model, on Friday. The model delivers significant gains in coding, scientific research, and cybersecurity while strengthening safeguards against misus…
A commentator observes that discussions about AI risk often assume adversaries like Trump and China would not cooperate to mitigate the threat, despite the existential danger to all parties. The author questions why thes…
Anthropic is negotiating with the Trump administration to restore global access to its Fable 5 and Mythos 5 AI models after a Commerce Department directive suspended access over national security concerns. The restrictio…
Autonomous vehicle development faces setbacks including a Tesla crash and Waymo recall, but the industry is doubling down with AI-driven partnerships. Stellantis, Uber, and Wayve plan Level 4 driverless taxis, while McKi…
The US Commerce Department has sent letters to Anthropic and OpenAI imposing ad hoc restrictions on AI model access and rollout, citing national security concerns, without formal legislation or regulatory framework. Anth…
A developer built a deterministic firewall for AI agents with cloud credentials, blocking destructive actions like network scans and escalating expensive provisioning calls to humans. The system uses zero-LLM checks to p…
A developer released a book titled 'Harness Engineering' on LeanPub, offering a pay-what-you-want model starting at $10. The book covers a discipline for building reliable workflows around non-deterministic AI agents, in…