AI Safety news and analysis on Web Pulse: 7774 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Researchers found that formal robustness certificates for embedded neural-interface models can pass even when task accuracy collapses, with EEGNet classification accuracy dropping 25.7% under attack while the certificate…
Researchers introduced Gimitest, an open-source framework for testing reinforcement learning policies across diverse environments and algorithms. The tool supports single- and multi-agent RL policies, aiming to improve r…
Researchers at arXiv show that biased LLM judges silently disable skill retirement in self-evolving agents, causing a mechanism failure that no amount of data can fix. The study reveals that false-pass bias, where failur…
Researchers introduced GhostWriter, a novel attack that poisons the long-term memory of personal AI agents by injecting hidden payloads, achieving 98% injection and 60% activation rates against state-of-the-art agents. T…
A comprehensive survey of 1,250 arXiv papers (2024-2026) on AI self-improvement reveals a taxonomy distinguishing bounded self-refinement from open-ended recursive self-improvement (RSI), with the latter constrained by g…
Researchers introduced institutional red-teaming, a methodology for testing deployment rules in multi-agent AI systems, and found that changing only the consequence rule shifts mean fatality by 22 to 58 percentage points…
Researchers propose PRoVeFL, a federated learning framework that combines privacy preservation, Byzantine robustness, and verifiable aggregation using multi-key fully homomorphic encryption across multiple servers. The f…
A developer proposes making AI-generated specifications compilable and executable to eliminate hallucinations in AI-assisted programming. By using a deterministic compiler instead of probabilistic AI to generate code, th…
OpenAI launched Patch the Planet on June 22, 2026, deploying AI agents to audit and patch open-source software, finding hundreds of vulnerabilities in the first week including four dnsmasq CVEs, a 23-year-old OpenBSD ker…
A developer proposes an AI agent autonomy ladder that maps tasks to five levels of autonomy—read-only, draft, copilot, supervised autopilot, and full autopilot—based on risk. The framework allows agents to act automatica…
A developer introduced basou, an open-source harness for steering AI coding agents. The tool addresses three pain points: the loss of decision context between sessions, the cost of re-explaining context, and the inabilit…
OpenAI, Trail of Bits, HackerOne, and Calif launched Patch the Planet on June 22, an open-source security initiative using AI-assisted vulnerability research with mandatory human review. The first five-day sprint covered…
Anthropic, GitHub, and Google all scheduled price increases for September 1st, with Anthropic raising Sonnet 5 prices after an introductory period and redeploying Fable 5 under tighter safety classifiers that cause more …
Independent lab Eigen Labs cracked Google's hidden quantum cryptography research in three days using crowdsourcing and AI agents, surpassing the original results. The breakthrough demonstrates the urgent need to migrate …
Oomol Lab released OpenConnector, an open-source project that provides a runtime boundary for SaaS access in AI agents, addressing issues like user account management, scope enforcement, and schema consistency across pro…
A developer built a Playwright script to test an LLM-powered customer support assistant's long-term memory and discovered three critical framework-level bugs. The automated test revealed that the memory module, built wit…
The U.S. Commerce Department ordered Anthropic to shut down its most powerful AI model, Claude Fable 5, to all foreign nationals, forcing the company to turn it off for all users worldwide just three days after its relea…
Researchers solved BlueDot's TAIS Puzzle #1 by identifying a nonlinear representation of the 'country' feature in a five-layer MLP, hidden as an XOR with 'food' at layer h2. They used Distributed Alignment Search to prov…
Anthropic's research reveals that language models can only verbally report about 10% of their internal activations, confined to a 'J-space' mental workspace. A new experiment using Natural Language Autoencoders (NLAs) on…
Researchers from AE Studio and Anthropic introduced Gradient Routed Auxiliary Modules (GRAM), a method that isolates dangerous knowledge to specific modules within a language model, enabling access control by switching m…