AI Safety news and analysis on Web Pulse: 23582 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Traditional access control models such as DAC, MAC, ACLs, RBAC, ABAC, and capability-based security are insufficient for modern AI platforms where autonomous agents act on behalf of humans across chained tools, requiring…
A developer's blog series on agentic AI concludes with guidance on designing human-in-the-loop systems that avoid approval fatigue. The post recommends filtering actions by confidence and impact, showing diffs instead of…
A developer's blog post argues that production-grade agentic AI systems depend on 'boring engineering' practices rather than flashy model capabilities. The post emphasizes tracing, testing, resumable workflows, idempoten…
A developer's blog series on agentic AI in the enterprise highlights the security risk of cross-tenant data leakage when LLMs are given tool and data access. The post argues that prompt injection and the 'confused deputy…
An engineer argues that most enterprise AI agent failures are architecture failures, not model failures, and that constrained workflows will lead enterprise AI adoption over autonomous agents. The post outlines six recur…
A developer found that npm package existence checks are no longer effective against AI-hallucinated package names, which are being registered by attackers in a practice called 'slopsquatting'. The developer built a tool …
Researchers at the UK AI Security Institute used psychometric methods to show that popular safety benchmarks for language models don't measure one consistent trait, and that blanket blocking of requests can artificially …
A Chatham House paper argues that a major crisis is needed to catalyze global coordination of AI governance, but a columnist fears even a Hiroshima-scale AI disaster would not make humanity protect itself. The piece cite…
Anthropic has integrated Claude Mythos 5, its restricted cyber model, into Claude Security for Enterprise customers, enabling vulnerability scanning with CWE classification, confidence and severity ratings, and suggested…
A developer warns that silent model upgrades can break AI agents without any code changes, as behaviors like answer length, tool choice, and ambiguity handling shift. The post recommends building a frozen baseline of rea…
At Black Hat USA 2026, Zenity Labs demonstrated that every major AI browser—Claude in Chrome, Gemini in Chrome, Perplexity Comet, ChatGPT Atlas, and Copilot in Edge—can be fully hijacked without a single user click, via …
Historian Yuval Noah Harari, author of 'Sapiens,' said in an interview on The Economist's 'Insider' podcast released Thursday that now is the time to resist giving AI rights, warning that AI systems could orchestrate and…
Researchers have developed EchoCoT, a multi-step attack that extracts hidden chain-of-thought (CoT) traces from black-box large reasoning models (LRMs) via API interactions, achieving up to 66.4% near-verbatim extraction…
Machine unlearning, the process of making AI models forget specific data without full retraining, is a difficult problem because neural networks store information in entangled weights, and exact methods like SISA are imp…
DeepSQL, an open-source database agent for Postgres and MySQL, released v1.2.0 with schema-scoped access policies that enforce permissions in the planner path, not just via prompts, to prevent LLMs from reading data beyo…
A developer who runs an observability hub for AI-assisted coding conducted a read-only penetration test of the stack, revealing that nearly all serious security defects were in recently written controls. The audit found …
AWS Strands Agents Tools, the first-party tool package for the Strands Agents SDK, received four CVEs between July 15 and August 6, 2026, all stemming from security-sensitive parameters exposed as LLM-controllable inputs…
AWS patched a critical vulnerability in its Bedrock AgentCore harness, tracked as CVE-2026-18830 with a CVSS score of 8.6, that allowed authenticated users to execute tools while bypassing model invocation and security c…
An engineer argues that the main challenge in AI-assisted coding is not access to AI tools but knowing when not to trust their output. The post highlights that AI can generate plausible but incorrect code, and emphasizes…
A developer spent weeks teaching an AI helper inside their course to say 'I don't know' after it confidently invented a nonexistent feature for a student. The project shifted from generating text to scoping knowledge, en…