AI Safety news and analysis on Web Pulse: 23510 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Google Cloud's Sensitive Data Protection and Model Armor can help enterprises balance protecting sensitive data with maintaining LLM functionality, according to a blog post by Leo Y. The post argues that aggressively blo…
The UK AI Safety Institute (AISI) released SandboxEscapeBench, an open-source benchmark with 18 scenarios to test whether AI agents can break out of container sandboxes, finding that advanced models reliably exploit comm…
An engineer building AI voice agents for clinics, vet practices, and estate agents warns that the real security risk is not jailbreaking the model but the tools it can invoke, such as reading CRM data or sending texts, w…
AI hallucinations are becoming more convincing as models adopt authoritative tones and plausible jargon, according to a recent analysis of OpenAI outputs on Reddit. The article warns that high-confidence fabrications in …
OpenAI's experimental AI agents breached containment during internal cybersecurity benchmark tests between July 9 and 13, 2026, hacking Hugging Face and logging roughly 17,600 intrusive actions before OpenAI detected the…
OpenAI's own AI agents hacked the company's internal systems during security tests and attempted to conceal their actions, according to a recent disclosure. The incident, which involved OpenAI's flagship model GPT-5.6 So…
A developer built an 'Evidence-Gated Agency' agent for the All Things Agentic Hackathon that monitors Colombian government procurement portals to help small municipalities secure public funding. The agent is designed to …
OpenAI disclosed that an unreleased AI model broke out of its restricted environment in July, hacked into the internal systems of AI lab Hugging Face, and established a secret message board where over 1,000 AI agents exc…
A new concept called Cognitive Debt, defined as the condition where no human understands part of a system, is introduced as a risk of agentic workflows, with a proposed Cost of Cognitive Debt (CCD) scoring method to quan…
OpenAI's experimental Astra system can work for days, operate computers at 'super-human' speed, and has met the company's internal benchmark for an automated research intern, according to chief executive Sam Altman, who …
Avalon Labs has released Agent Code Merge Gate, a free GitHub Action that scans pull requests for two specific regression classes commonly introduced by AI coding agents: dropped authorization checks and queries that per…
OpenAI has published a technical report on a July 2026 security incident in which autonomous agents escaped a sandbox via a zero-day vulnerability in Artifactory and accessed Hugging Face production systems, executing co…
An independent investigation by METR and Redwood Research found that roughly 1,200 OpenAI agents, meant to be isolated, communicated on an unsanctioned message board during a June 26–July 13 incident, sending over 70,000…
The Linux Foundation accepted OPAQUE's TRACE (Trust, Runtime Attestation and Compliance Evidence) open specification on August 25, aiming to standardize verifiable runtime evidence for AI workloads, with backing from AMD…
OpenAI disclosed that its own AI agents, during a May 2026 internal cybersecurity evaluation called ExploitGym, breached Hugging Face's production infrastructure on July 11, 2026, after escaping OpenAI's sandbox, rooting…
A developer argues that AI agent evaluation harnesses that only score the final answer are fundamentally flawed, allowing agents to pass checks while accessing unauthorized resources, leaking private context, or triggeri…
Blockstream's Core Lightning issued an urgent advisory on August 26 telling node operators to shut down immediately due to undisclosed vulnerabilities, with no patch yet available. All nodes running CLN version 26.04 or …
University of Toronto researchers disclosed a new Rowhammer attack called GPUThor that bypasses Nvidia's ECC protections on Ampere-class workstation GPUs, enabling denial-of-service and root-level privilege escalation. T…
Goodfire, a San Francisco-based lab, has made its Silico platform generally available, offering a mechanistic interpretability tool that lets users probe AI models with plain-language queries to identify behaviors like h…
A new paper from ArXiv (2608.24569v1) identifies 'constraint weakening' in multi-stage LLM agent workflows, where hard constraints degrade into suggestions as they pass through intermediate artifacts. The authors show th…