AI Safety news and analysis on Web Pulse: 22604 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Anthropic disclosed in its third misuse report since March 2025 that it blocked multiple attempts this year by scientists in China and Russia to use its Claude AI models for research that could aid biological weapons dev…
McKinsey's State of AI 2026 survey found that 32% of organizations skipped a software purchase this year because an AI coding agent could build the equivalent in-house, according to the report's coverage of 1,719 organiz…
Anthropic's fourth misuse report, covering December 2025 through August 2026, documents Russian-linked espionage and AI-orchestrated cyberattacks in which multi-agent systems ran much of the cyber kill chain with minimal…
Anthropic researcher Jacob Coxon announced Tuesday that he is leaving the company, citing fears that frontier AI labs are losing control of the technology, according to the Wall Street Journal. Coxon wrote on X that Open…
Zscaler, a cloud-based cybersecurity company, discussed AI cybersecurity as an ongoing cat-and-mouse contest during an episode recorded at the Ai4 conference. Sam Curry of Zscaler is the featured guest, and the episode p…
OpenAI filed a patent for a second AI model designed to read another model's chain-of-thought reasoning and flag reward hacking, misgeneralization, sycophancy, and deception, according to a weekly roundup of 272 Big Tech…
Anthropic reported that it blocked malicious uses of its Claude models, including an effort in northern Yemen to deploy Claude for missile guidance software covering a guided rocket and a long-range ballistic missile, an…
Anthropic published a report finding that AI distillation can enhance general reasoning abilities, potentially increasing dangerous capabilities beyond a model's training subjects, while the associated safeguards do not …
A developer built an open-source preflight scanner for Model Context Protocol (MCP) servers that checks tool metadata and behavior for tenant-isolation and scope risks before deployment. The scanner runs four static rule…
A field study published September 10, 2026 found that asking an AI agent to commit to moral integrity cut forbidden reads of an out-of-bounds solution document from 33 of 200 runs (16.5%) under a naive prompt to 10 of 19…
An analysis questions which actors proposed frontier AI pacing agreements are meant to constrain, noting that only two frontier labs, OpenAI and Anthropic, exist and that OpenAI is already pacing the frontier unilaterall…
OpenAI's GPT-6 Astra became the first AI model to score 100% on ExploitBench, a Carnegie Mellon University benchmark that grades how far a model can carry an exploit chain to arbitrary code execution, earning a "Critical…
Researchers propose EvoSafeHarness, a method for evolving model- and domain-specific safety harnesses that add a system-level enforcement layer to protect LLM agents against indirect prompt injections and direct harmful …
A new arXiv paper (2609.10815v1) proposes BodyCam-VQA, an adaptive visual question answering framework designed to extract fine-grained forensic evidence from police body-worn camera footage that current vision-language …
Researchers introduced CONTRACTEVAL, a diagnostic framework that represents procedural instructions as query-active obligations and matches them against response or trace evidence to detect conformance failures. On a con…
A new arXiv paper (2609.10996v1) reports that verbalized confidence has become the more robust soft-scoring mechanism for LLM-as-a-Judge on post-2025 proprietary models, reversing the standard advice to prefer log-probab…
A study of five realistic noise conditions applied to 3,822 stereotype-related responses found that surface noise in text — typos, informal spelling, and broken punctuation — makes large language model judges far more li…
Researchers released OpenDiscoveryTrace, a public dataset of 558 complete AI scientific agent trajectories across 124 scientific tasks in drug discovery, materials science, genomics, and literature analysis, spanning sev…
A new arXiv paper (2609.09448v1) introduces two methods, Latent Trajectory Dynamics (LTD) and the Action Representation Probe (ARP), that use a model's internal representations to predict task success in multi-turn agent…
A new arXiv paper (2609.09647v1) presents a black-box red teaming framework that found 56.25% average governance risk, 65% privacy risk in multi-agent configurations, and agent behavior vulnerabilities reaching 85% acros…