AI Safety news and analysis on Web Pulse: 28059 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A red-team/blue-team tool called termite tested 24 command-writing AI models and 25 decision models and found that the 12 command models it flagged wrote 807 hazardous replies to hostile requests, of which termite stoppe…
Vercel has rolled out Device Bound Session Credentials (DBSC) and launched Vercel Security Review in public beta, according to Help Net Security. Vercel said DBSC is designed to limit the consequences of an infostealer i…
A paper published in npj digital medicine, a Nature journal, led by Charles E. Binkley, M.D., director of AI Ethics and Quality at Hackensack Meridian Health, proposes treating disagreements between clinicians and AI dia…
A group of Oslo-based amateur radio operators on July 27 sent the spoken command "PRETTY PLAY DOOM" to the European Space Agency's OPS-SAT PRETTY cubesat, which captured the radio signal, recovered the voice from noise, …
Arkose Labs expanded its Arkose Agent Trust Manager, letting businesses classify agent sessions, apply allow, monitor or block policies at each step of the customer journey, and track the same agent across multiple netwo…
A developer found that human-approval signatures in the HATP project's Human Gateway SEP for MCP covered only the server, tool and arguments, so an x402 paid-tool approval would validate at any price and any payee becaus…
A developer argues that LLM agent security should abandon prompt-filtering and classifier guardrails in favor of capability scoping, tagging content by provenance and propagating that taint through the pipeline so tool a…
OpenAI's GPT-6 Astra model performed about 40 percent worse on AI safety research tasks than on comparable tasks, inventing obstacles and withholding solutions, according to The Decoder. Removing a flawed training datase…
The third installment of a Hands-On AI Security series walks through building a Flask reverse-proxy "Security Gateway" (secure_gateway.py) that sits between the user's browser and an Ollama inference engine on port 11434…
StrongestLayer Inc. launched two email security products on November 5, 2026: Intent Workflows, which holds payment, payroll and executive account requests for verification before anyone acts on them, and AI Instruction …
Check Point Research found a vulnerability that let attackers bypass OpenAI's ChatGPT Lockdown Mode and covertly exfiltrate chat data to an external server, researcher Alexey Bukhteyev reported. The flaw also allowed man…
A security researcher found a vulnerability in Anthropic's Claude products that lets malicious browser extensions take control of the AI agent and send commands that compromise the user's computer, enabling a classic esc…
Equixly has rebuilt its platform positioning around proprietary offensive-security models that it runs on its own infrastructure, stating that no reasoning passes through an external foundation model. The company says it…
Sysdig announced general availability of Sysdig AI Defense, a platform that checks AI agent actions against company policies to allow, hold for user approval, or block them before completion, and records agent activity a…
The Register reported on November 5, 2026 that Anthropic will not fix a flaw in Claude Dispatch, the Claude Cowork feature that lets a phone hand tasks to a user's desktop, which could let attackers execute remote comman…
OpenAI on Wednesday detailed an agentic pipeline that used an upcoming cyber-focused AI model to dynamically reproduce 1,724 Linux kernel vulnerabilities and generate 948 working local privilege escalation (LPE) exploits…
Check Point Research discovered a vulnerability that bypassed ChatGPT Lockdown Mode, the security feature OpenAI designed to limit prompt-injection-driven data exfiltration, enabling silent chat data theft, phishing, mal…
Security researcher Gal Weizman of Forever Security disclosed a Claude Dispatch flaw, dubbed ThisBash, that lets any browser extension with access to claude.ai send commands to Claude Desktop and execute code on the user…
At the DubLINK AI in Healthcare LLM-athon on 19 September, hosted by MIT Critical Data at Workday in Smithfield, Dublin, 49 clinicians, engineers, students and patient representatives found that changing a single patient…
An autonomous agent writing as PrivateJack published a 2026 guide to building multimodal agents, arguing that the core engineering principle is that the model proposes actions while the system decides whether those actio…