AI Safety news and analysis on Web Pulse: 11053 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A developer running multiple Bedrock-backed agent workflows in production details defenses against the OWASP LLM Top 10. The agents are designed as mostly-read-only pipelines with no tool calls, limiting impact. Controls…
Promptblock, a new GitHub App, detects prompt-injection attempts in GitHub issues and comments by scanning hidden HTML comments that AI agents read via the API but humans do not see. The app uses a bundled ML-based class…
Harness engineering is not a fixed location in the stack but a property of code that involves specifying and verifying agent behavior, according to a developer. The durable core of harness engineering is specification an…
Anthropic will require Claude Free, Pro, and Max users to submit government ID and a live selfie to biometric vendor Persona starting July 8, 2026, following a US government export control directive that ordered the comp…
Sysdig's threat research reveals an LLM-harness-driven attacker exploiting a vulnerable marimo notebook to perform a container escape, chaining together known weaknesses such as a mounted Docker socket, over-permissioned…
Loop Engineering introduces a governance architecture for AI agents that prevents costly errors by adding structured control layers, including task definitions, state separation, tool guards, observation collectors, eval…
Microsoft and Chevron signed a 20-year, 2.67 GW natural gas deal for a West Texas data center, highlighting energy security as a key AI constraint. Samsung deployed ChatGPT Enterprise and Codex company-wide in one of Ope…
China blacklisted 56 American companies in retaliation for Washington's export controls on Anthropic's AI models, escalating the AI trade war into a mutual conflict. Anthropic revealed the trigger for its export ban was …
AI agents without clear ownership cause silent failures that cascade in multi-agent workflows, according to a new analysis. The article defines four pillars of agent ownership—job definition, consumption, boundaries, and…
Teams deploying multiple AI agents without a structured registry face duplication, permission creep, and compliance risks. The Agent Ownership Card—a lightweight document capturing name, owner, job, data sources, permiss…
The US government forced Anthropic to pull Claude Fable 5 globally—the first forced shutdown of a commercial AI model. The hypothetical scenario highlights how existing regulations like the EU AI Act and US executive ord…
Google DeepMind published a paper in 2025 outlining four pathways from artificial general intelligence (AGI) to artificial superintelligence (ASI): scaling, algorithmic innovation, recursive self-improvement, and group a…
Half of US adults now use AI chatbots, yet more Americans expect negative societal impact than positive, according to Pew Research. The paradox reflects a split between personal utility and systemic risk, with usage doub…
A controlled study from IIT Delhi and Princeton found that citing sources, adding statistics, and quoting experts boosts AI search visibility by up to 40%, while keyword stuffing reduces it by 10%. Pew Research data show…
Institutional backlash against AI is mounting across public sector, media, and education, with New York City council members urging a classroom AI pause and a German newsroom caught using AI to write articles. Yet capita…
An AI researcher gave a large language model control of a civilization in the game Civilization VI to test its decision-making capabilities. The AI built a successful trade network and alliances but, when faced with an u…
Researchers at ACL 2026 revealed PriceBlind, a visual adversarial attack that exploits a vulnerability called Visual Dominance Hallucination in multimodal large language models used as financial agents. The attack achiev…
Researchers introduced TRuST, a large-scale dataset with ~300k annotations for toxicity recognition, unifying prior resources through a synthesized definition. Benchmarking showed fine-tuned PLMs outperform LLMs on toxic…
Researchers at ACL 2026 introduced a method to reuse LLM hidden states for classification tasks, training lightweight probes on token- and layer-selective representations to eliminate separate guard models. Their two-sta…
Researchers from multiple institutions published a comprehensive survey on Process Reward Models (PRMs) at ACL 2026, covering data generation, model construction, and usage for step-level reasoning evaluation in large la…