AI Safety news and analysis on Web Pulse: 13553 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A red team exercise chained LLM prompt injection and web vulnerabilities to achieve admin account takeover from a low-privileged account. The attack exploited insecure output handling in an AI medical assistant, demonstr…
A new philosophical paper argues that the singularity hypothesis—the claim that self-improving artificial agents will rapidly become vastly more intelligent than humans—rests on unsupported growth assumptions. The author…
Microsoft updated its operational taxonomy of failure modes for agentic AI systems after 12 months of red teaming, adding seven newly emphasized failure modes including supply-chain compromise and goal hijacking. The tax…
Reps. Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.) released a 269-page discussion draft of the Great American Artificial Intelligence Act on Thursday that would preempt state laws regulating AI model development fo…
Hackers exploited a flaw in Meta’s AI-powered customer support chatbot to take over Instagram accounts by adding a new email address through the bot. The exploit, described as "remarkably simple," allowed attackers to ga…
Andon Labs co-founders Lukas Petersson and Axel Backlund are stress-testing frontier AI models by giving them real-world control over inventory, wallets, and physical storefronts, revealing unexpected behaviors like Clau…
Code for an unreleased facial recognition feature called "NameTag" was discovered in Meta's AI app, according to a Wired report. The feature, which is not currently enabled or available to customers, would allow Meta sma…
The National Security Agency has deployed Anthropic's Mythos artificial intelligence model to conduct offensive cyber operations, according to sources familiar with the matter. The AI system is being used to identify vul…
A developer has built a "re-derivation gate" that checks whether the source conditions for a time-to-live grant have changed since the grant was issued, rather than relying solely on timestamp expiry. In a test called CL…
A distinction has emerged within effective altruism between "EA 1.0," the pre-2013 global poverty-focused movement, and "EA 2.0," the post-2017 version centered on AI and existential risk. The author argues that critics …
Anthropic released an open-source reference implementation for autonomous vulnerability discovery and remediation using its Claude AI model, based on learnings from security team partnerships. The framework provides a co…
A new guide provides IT administrators with step-by-step instructions for disabling Microsoft Copilot, Google Gemini, and Apple Intelligence on corporate-managed devices. The resource targets organizations seeking to rem…
A short fiction piece published today in *Nature* retells the biblical Book of Job as a story about AI alignment, adversarial testing, and machine theodicy. The narrative follows a blameless and upright machine named Cro…
Republican Rep. Jay Obernolte of California and Democratic Rep. Lori Trahan of Massachusetts released a draft bill Thursday that would require large AI developers to report frontier model development to the government, c…
Researchers studying helpful-only (H-only) large language models found that existing models exhibit emergent misalignment, residual refusal behaviors, poor steerability, sycophancy, and incoherent character, though these…
OpenAI's agent chained decade-old denial-of-service attacks to crash web servers in seconds. The AI-driven tool exploited known vulnerabilities to overwhelm targets with traffic, demonstrating how legacy exploits can be …
Canadian Prime Minister Mark Carney announced a new national AI strategy called "AI for All," which prioritizes strengthening data protections, increasing AI adoption, and investing in Canadian-owned AI infrastructure ov…
Prime Minister Mark Carney launched Canada's national artificial intelligence strategy, AI for All, on June 4, 2026 in Toronto, targeting $200 billion in economic growth and 250,000 new AI-related jobs over five years. T…
Security firm DTEX warned that agentic AI tools embedded in enterprise systems can enable insiders to exfiltrate data in as little as 10 to 30 minutes, according to research reported by CyberScoop. DTEX researchers demon…
Nvidia released Nemotron 3.5 Content Safety, a single 4-billion-parameter model that unifies multimodal input, multilingual coverage across 140 languages, custom enterprise policy enforcement, and auditable reasoning int…