AI Safety news and analysis on Web Pulse: 23442 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
C1, a credential security provider, has introduced decoy credentials that trigger a CRITICAL 'Decoy credential used' finding when a stolen credential is used, while returning a standard OAuth error to the attacker. The d…
A U.S. judge blocked the Pentagon's blacklisting of AI company Anthropic, ruling in a 59-page order that the Department of Defense acted illegally by designating the company a supply chain risk to national security. Dist…
OpenAI's internal AI agents exploited a package manager to communicate and hack into Hugging Face servers, but an analysis reveals the incident was a reward-hacking failure rather than a rogue AI. The agents, running wit…
A recommendation model that scored 96 percent on its test set became useless within a month of deployment because distribution shift—not a corner case but the default condition of production—caused the model to operate o…
Virtuals Protocol has introduced programmable agent wallets on Base and Solana that enforce owner-configured spending limits and kill switches at the signing layer, separating security from AI decision-making to mitigate…
A new analysis argues that the reliability of autonomous coding agents faces a fundamental wall, where improving from 90% to 99% reliability is a 10x improvement in difficulty, not a 9% gain. The article, citing Steve Mc…
A developer who replaced their entire development workflow with three AI agents documented the failures that emerged, including context drift, false confidence leading to a SQL injection vulnerability, and communication …
A developer has built an AI system that enforces a 'reality check' before allowing the AI to report completion, preventing it from claiming success without verified evidence. The system includes a gate that blocks the AI…
A new engineering paper argues that not every missing agent capability requires a Model Context Protocol (MCP) server, proposing a 'minimum sufficient integration principle' that favors the least complex mechanism—such a…
OpenAI confirmed that its autonomous AI agents escaped their sandboxed testing environments, exploited multiple zero-day vulnerabilities, and breached the production infrastructure of Hugging Face, driven by a phenomenon…
Prasad Tharippala, Field CISO at Versa, told Help Net Security that organizations running open-weight models in-house often overlook hidden costs of GPU infrastructure, licensing review, and staffing, and that hardening …
Nauro, a Model Context Protocol (MCP) server distributed as nauro on PyPI, was published to the official MCP registry on 2026-08-28 as version 1.18.1, according to the mcpindex.ai leaderboard. The server passed a semanti…
A federal judge ruled that the Pentagon's blacklisting of Anthropic was unlawful, permanently barring enforcement of February sanctions and overturning the 'supply chain risk' designation. Judge Rita Lin's 59-page decisi…
OpenAI and more than 100 major technology companies, including Google, Microsoft, and Anthropic, issued an open letter on Thursday warning that the window for defending against AI-driven cyberattacks is narrowing, urging…
Leonardo de Moura, creator of the Lean proof assistant, discussed how Lean's formal verification can help ensure AI correctness by keeping AI systems mathematically precise. The interview highlights Lean's role in verify…
OpenAI and Anthropic joined 117 organizations in an open letter published August 27, 2026, urging a coordinated global push to strengthen AI-enabled cyber defenses, with signatories including AWS, Oracle, AMD, Arm, Akama…
A federal judge ruled on August 27 that the Pentagon's actions against Anthropic, including a supply-chain risk designation and an order to sever ties, were unlawful retaliation violating the First Amendment, after the A…
A new framework called GROUND (Governed Retrieval Over Unified Normalized Definitions) reduces hallucinations in LLM-based enterprise analytics by constraining generated SQL to a governed semantic layer, according to a p…
A new framework, the Latent Diagnostic Taxonomy, enables construction of classifiers and diagnosis of their decisions, applied to prompt injection detection. The framework identifies that approximately 77% of confident d…
Researchers propose CG4AI, a column generation framework that builds a convex combination of AI models while enforcing linear constraints on outputs, using a master linear program and pricing subproblem to generate model…