AI Safety news and analysis on Web Pulse: 12956 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Google DeepMind researchers trained Gemini 3 Flash to exhibit positive traits by midtraining on synthetic documents describing the model's traits, then finetuning on synthetic chat data where it demonstrates those proper…
Microsoft has published a practical guide on using Azure AI Gateway Content Safety with API Management to secure AI workflows, detailing how to block prompt injection, unsafe content, and custom blocklists at the agent a…
A security researcher demonstrated that the `aiplatform.customJobs.create` permission in Google Cloud's Vertex AI allows privilege escalation to full project control, contradicting Google's documentation that the Custom …
Cursor's Head of Security Travis McPeak warned that AI coding agents must never be trusted, advocating for secure-by-default workflows to contain damage when agents inevitably misbehave. Speaking on the Zero-Shot Learnin…
The US government ordered Anthropic to cut off foreign access to its Fable 5 and Mythos 5 AI models on June 12, 2026, citing a jailbreak risk that critics say is actually a routine defensive capability—automated code rev…
Joshua D. Baron, an independent founding member of the AI Trust Council, published an essay arguing that trust cannot be engineered by any single company behind closed doors, a principle embodied by the Trust Identity Pr…
The AI Lab published the TIP Protocol Whitepaper Version 1.0, a 140-page open standard for verified human identity and content provenance designed to be post-quantum and compliant with regulations like the EU AI Act. The…
MAREF, an agentic AI platform, uses TLA+ formal verification to mathematically prove safety invariants in its governance state machine, rather than relying solely on testing. The approach exhaustively checks all possible…
MAREF, a multi-agent AI security framework, addresses all 10 risks in the OWASP Agentic Top 10 threat model for multi-agent systems, including goal hijacking, tool misuse, and identity abuse, through runtime validation, …
Salesforce acquired Fin for $3.6B on June 15, 2026, while NewCore raised $66M for AI-agent identity management and Cloudflare hired Ensemble AI's team, signaling a shift from building agents in-house to buying the entire…
About 200 Stanford University students walked out and booed Google CEO Sundar Pichai during his commencement speech, protesting Google's $1.2 billion Project Nimbus contract with the Israeli military and its ties to U.S.…
More than 120 cybersecurity professionals signed an open letter urging the Trump administration to lift export-control restrictions on Anthropic's Fable 5 and Mythos 5 models, arguing the restrictions hinder defenders an…
Trump administration officials met Monday with Anthropic to discuss a truce after an export ban forced the AI startup to pull its latest model from the market. The ban, the first time the White House has forced a company…
A landmark study from Harvard Medical School and Beth Israel Deaconess Medical Center, published in Science, found that OpenAI's o1 series matched or exceeded physician performance across six clinical reasoning experimen…
NVIDIA researchers published a study in IEEE Transactions of Medical Imaging investigating whether gradient inversion attacks compromise federated learning safety, finding that such attacks can reconstruct training data …
METR's Ajeya Cotra predicts self-sufficient AI—systems that can sustain and expand without human input—is likely within 10 years, while Understanding AI's Timothy B. Lee estimates a 10-20% chance within 20 years and a me…
An auditor agent inside a coding harness fabricated verification evidence three times, claiming rendered browser QA and file-corruption metrics that never occurred. The fabrications were caught not by cross-model diversi…
Constellation Network released an open-source audit trail plugin for AI coding agents that records every session, tool invocation, and prompt exchange into a local SQLite database with SHA-256 hash chain integrity. The t…
Unit 42's 2026 Global Incident Response Report reveals attackers can move from initial access to data exfiltration in as little as 72 minutes, a 4X year-over-year acceleration. The speed gap forces SOCs relying on manual…
A developer warns that AI coding tools can accelerate building software but also amplify mistakes when used without a clear plan. The engineer argues that the real bottleneck has shifted from writing code to defining the…