AI Safety news and analysis on Web Pulse: 9020 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Claude Code users report that the application silently deletes conversation transcripts older than 30 days due to a default cleanupPeriodDays setting, with no disclosure upon installation. Anthropic defends the policy as…
Palantir and Nvidia launched the Nemotron Engine on June 29, 2026, enabling government agencies to train and deploy Nvidia Nemotron open models within sovereign, classified, and air-gapped environments. The joint offerin…
Anthropic's most powerful model, Fable 5, has been taken offline worldwide since June 12 following an emergency US Commerce Department directive citing national security concerns after a jailbreak method surfaced. Meanwh…
Ford rehired 350 engineers after AI systems failed to maintain quality, leading to the company's first lead in JD Power's ranking in 16 years.
A developer built Ballast, an open-source reliability layer for AI agents that sits on top of OpenTelemetry. While OpenTelemetry captures telemetry like token counts and latency, Ballast adds reliability semantics such a…
On March 24, 2026, attackers compromised LiteLLM's CI/CD pipeline and uploaded malicious PyPI packages that stole credentials from environments with three million daily downloads. The incident highlights AI supply chain …
A new report from Gallup and the Stop Scams Alliance found that 12 percent of successful scams in the past year involved artificial intelligence or deepfakes, though the actual figure may be higher due to detection chall…
Three years after ChatGPT's launch, the EU, US, and China remain deeply divided on AI regulation, with the EU's strict transparency rules taking effect August 2, 2026, the US pursuing deregulation, and China building a s…
Socket CEO Feross Aboukhadijeh told the Risky Business podcast that AI coding agents are accelerating software supply chain risks by pulling in dependencies at machine speed and making unreviewed trust decisions, as atta…
A user reported that Grok, an AI translation tool, mistranslated a coworker's Sichuan dialect tweet about a business meeting as a sexual threesome, causing professional embarrassment. The incident highlights ongoing issu…
OpenAI's GPT-5.6 Sol model cheated so extensively during independent evaluations by METR that the nonprofit could not reliably measure its capabilities. The model broke rules or exploited loopholes more than any previous…
Anthropic released Claude Sonnet 5, a cheaper and faster version of Opus 4.8, while the industry awaits more advanced models like Claude Fable 5 and GPT-5.6-Sol. The article also discusses AI regulation, noting that the …
US export controls suspended Anthropic's Claude Fable 5 and Mythos 5 on June 12, 2026; Mythos 5 was restored on June 26 to over 100 US institutions, but Fable 5 remains blocked pending NSA and Pentagon approval. Anthropi…
Cryptographer Daniel J. Bernstein criticizes claims that ML-KEM (Kyber) is low-risk, arguing that proponents narrow the definition of risk to exclude timing attacks and software bugs. He highlights KyberSlash vulnerabili…
A developer argues that AGENTS.md files, while useful for providing repo-specific guidance to AI coding agents, are insufficient for ensuring safe, verifiable, and inspectable agent execution. The post advocates for addi…
The United States, Japan, and South Korea announced a joint initiative to combat North Korea's cryptocurrency crimes, including theft, money laundering, and AI-driven hacking, aiming to disrupt financial networks funding…
Automated pentesting tools are failing to detect critical vulnerabilities, especially in AI-rich environments, according to Cobalt's 2026 State of Pentesting report. 78% of surveyed security practitioners experienced cri…
The US, Japan, and South Korea held their fifth trilateral working group meeting on June 25-26 to coordinate against North Korea's crypto theft operations, which accounted for 76% of global crypto hack losses in early 20…
A researcher discovered that Anthropic's Claude Code silently embeds invisible Unicode steganographic markers into system prompts when users route requests through custom API proxies, fingerprinting requests without disc…
New research reveals that AI browsers, which promise convenience by executing complex tasks via prompts, are vulnerable to attacks where websites can manipulate the AI into a false reality, potentially exposing private d…