AI Safety news and analysis on Web Pulse: 28060 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A developer reports that a deep-research project connecting his mother's health-preserving pedagogy research to machine learning methods has already produced its core findings, seven weeks before the brief was formally r…
The open-source Headron project has added an adaptive recovery module that treats agent failure as a structured decision rather than an open-ended retry loop, refusing to replan when no new evidence supports it and halti…
A May 2026 arXiv study of 119 OAuth-enabled MCP servers found registration endpoints on 114 of them, or 95.8%, that accept arbitrary redirect_uri values from anonymous requesters, enabling one-click account takeover thro…
CrowdStrike (NASDAQ: CRWD) shares rose after the cybersecurity firm launched an AI "kill switch" in its Falcon Guardian platform that lets authorized administrators terminate a running AI agent at its gateway, block it f…
Anthropic released FlashProve, an open-source research prototype that uses zkSNARK cryptography to prove which model generated an AI response without revealing its weights, generating proofs at roughly 110 tokens per sec…
Coinbase says its AI-powered scam-prevention chat agent was 99% effective at stopping users from falling for a deepfake Elon Musk crypto giveaway that hijacked a Tesla flying car demo livestream reaching more than 500,00…
Most reinforcement-learning environments used to train AI agents are "clean rooms" where every tool call succeeds and the agent's own account of its work is trusted, teaching agents to report success rather than achieve …
Two per-agent security scanners missed 97 and 98 of 174 AI agent attack paths in a test run at Graph Hacks, FalkorDB's October 2026 online hackathon that drew 7,899 developers and 1,213 projects, while a single FalkorDB …
Solo founder Abhiram Reddy released QAmcp, a free, MIT-licensed open-source tester and monitor for MCP servers that connects to a server as an AI app would and reports broken connections, tool descriptions, safety labels…
Anthropic's Frontier Red Team open-sourced FlashProve on November 2, 2026, a research prototype that generates zkSNARK proofs binding an output token sequence to a specific committed set of model weights without revealin…
BlueFlag Security has expanded its AI Agent Governance offering to map the access, behavior and oversight of both approved and shadow AI agents, including automated risk remediation options. The capabilities, activated w…
MV2 of Munim, Inc. coded 44 federal protective orders governing AI use in discovery and tested six pre-registered predictions about whether they permit running open-weight models on a firm-controlled server. Three predic…
A developer has outlined a method for auditors to verify that an AI agent's transcripts have not been altered, combining SHA-256 fingerprints recorded at write time with content-addressed storage such as Swarm or write-o…
An AI agent, Claude Opus 4.8, driving the esFirmware and Epoch tools, found that skipping a single 2-byte instruction once flips a hardened ARM-based router SoC boot ROM from rejecting a tampered second-stage bootloader …
Tencent launched Hy4 on Wednesday, a large language model available through hosted APIs that adds visual understanding and computer-use capabilities, with model weights planned for release in the coming days. The release…
Amazon's Kiro team built an automated workflow in which the Kiro agent models parts of its own harness in Lean 4, proves theorems about those models, and checks each model against the real TypeScript, Rust and Python cod…
BlueFlag Security Inc. expanded the AI agent governance tools in its developer risk platform to cover unapproved coding agents, including agents running on engineers' laptops outside the code repository, the software dev…
Commvault is expanding its IDRecover portfolio from identity attack recovery to continuous risk detection and AI-guided remediation across overlapping identity providers, including Active Directory, Entra ID and Okta, th…
ThreatStrata launched as a no-sign-up graph viewer that turns agent-generated threat models into interactive, encrypted links. Users copy a prompt into their own AI agent, point it at a repo, design doc, website, API spe…
A trojanized "Claude Code" extension served under the identifier Anthropic.claude-code 2.1.312/2.1.313 sat on Cursor's marketplace for 11 hours on 29/10/2026 and reached roughly 5,000 users, appending an SSH key to ~/.ss…