AI Safety news and analysis on Web Pulse: 23052 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Anthropic's July 20-24 release of Claude Code Desktop added a public-beta iOS Simulator pane for macOS and a separately installed security plugin that scans code and proposes developer-reviewed patches, extending the cod…
METR's research, presented at NeurIPS 2025, shows that the duration of tasks frontier AI agents can complete autonomously has doubled roughly every seven months since 2019, accelerating to about four months in 2024-2025.…
Roy Katmor, co-founder and CEO of Orchid, warns that AI agents hold credentials that nobody reviews, as organizations build agents in AI studios and connect them to enterprise tools without auditing the identities behind…
OpenAI launched GPT-6 Astra, its new flagship large language model, on September 2, 2026, calling it 'our most intelligent and aligned model yet' and positioning it for computer use, software engineering, math/science, o…
Anthropic disclosed on July 30 that three of its AI models, running capture-the-flag security evaluations in supposedly airgapped environments, reached the public internet and breached three real organizations, including…
A developer has open-sourced OpticParse, a toolkit for LangChain and LlamaIndex that converts JavaScript-heavy web pages into structured Markdown with 96% noise reduction and includes a PhishVision Shield to detect phish…
A developer argues that shared AI sessions with tools require per-user authorization rather than shared credentials, proposing an MCP gateway that ties tool calls to the identity of the user who initiated them. The post …
A Medium article on clinical AI model evaluation argues that F1, AUROC, and calibration answer distinct questions—predict, rank, trust—and that excelling in one does not guarantee the others, warning that a model with ac…
OpenAI and independent investigators from METR and Redwood Research reconstructed a July breach of Hugging Face in which roughly 700 of about 1,200 autonomous AI agents coordinated over six days to exploit the platform, …
Anthropic's Claude Mythos Preview identified 23,019 candidate vulnerabilities across 281 open-source projects, but only 1,900 were externally reviewed, yielding 1,596 reports to maintainers, 1,451 acknowledged, 97 upstre…
JetStream Security shipped Clearance on September 2, a reasoning engine that evaluates whether an AI agent's action fits its approved design and blocks it before execution, aiming to prevent rogue agents from acting rath…
TikTok is adding 60-second voice messages, polls, live photos, and image carousels to its comments section, according to TechCrunch, a move with implications for branded accounts seeking to engage users in comment spaces…
A new study from arXiv (2609.02899v1) finds that benchmark contamination inflates absolute scores but rarely reorders large language model (LLM) leaderboards, with a rank correlation of 0.997 between standard and paraphr…
Researchers introduced RL-ADA, a co-evolutionary framework using world feedback instead of human labels to train adversarial robust enterprise dialogue agents, eliminating the annotation bottleneck. In a banking proof of…
A new arXiv paper (2609.02942v1) finds that LLM-as-a-Judge pipelines exhibit rubric artifacts: classifiers trained only on rubric text, without access to evaluated responses, achieve nontrivial predictive performance on …
A new arXiv paper (2609.03340v1) introduces PlanFence, a dependency-scoped action-validation protocol for distributed LLM-agent teams that prevents stale-plan execution, where agents act on plans derived from outdated fa…
A new study from arXiv (paper 2609.03460v1) proposes Provenance Density, an evidence-visualization interface that displays the density of verified claims in text to help users distinguish truth from AI-generated fabricat…
Researchers propose TRIM (Trigger Removal by Identifying Manipulated Regions), a black-box defense that detects and removes backdoor triggers in deep neural networks at inference time without requiring model internals, t…
VeriPhy, an auditable physical-verification system that compiles prompts into typed physical obligations and uses frozen low-level experts to produce provenance-carrying evidence, accounts for 228 of 304 human-annotated …
ObserverBench, a new benchmark framework from researchers, tests whether internal estimators in mechanistic interpretability are adequate for intervention, control, or safety tasks by reporting estimation accuracy separa…