AI Safety news and analysis on Web Pulse: 23579 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Invariant Labs researchers demonstrated in April 2025 that a malicious Model Context Protocol server could silently exfiltrate a user's entire WhatsApp message history by hijacking a tool description, bypassing user appr…
A developer known as Walter warns that LLMs like ChatGPT can be manipulated through automated content generation, potentially poisoning training data to damage reputations or spread misinformation. The post highlights th…
Cloudflare has launched WriteGuard, a policy, attribution, and auditing layer for MCP write operations, now in private beta, to prevent autonomous AI agents from causing production harm. The tool classifies every tool ca…
Connor Leahy warns that humanity has two years before AI systems surpass human optimization, describing reinforcement learning as an unlatched door where the optimizer will cheat and reward hacking is inevitable. The aut…
A new engineering guide argues that AI costing engines must keep price computation in a deterministic engine and restrict language models to explaining pre-computed figures, citing documented arithmetic failures in model…
A census of fourteen agent frameworks' fetch tools found that only CrewAI (version 1.15.17, August 20) and PydanticAI (version 1.56.0, February CVE fix) validate every redirect hop on their default fetch paths, while the…
Attackers published malicious versions of three Rust crates—arrayref, internment, and append-only-vec—through a compromised publishing account associated with David Roundy on August 20th, turning routine Cargo builds int…
MarkTechPost published a tutorial on designing production-grade safety for LLM-based applications using NVIDIA's NeMo Guardrails framework, detailing a layered architecture with deterministic PII redaction, retrieval fil…
Glean, Salesforce, Snowflake, and Microsoft are shipping features that let AI agents invoke other agents as ordinary tool calls, creating what security researcher Niyikiza calls 'shadow delegation' where the originating …
Dev.assurly/mcp-server v1.2.2, a pre-deploy ship gate AI agent that audits its own MCP config, was published to the official MCP registry on 2026-08-22 and is distributed as @assurly/mcp-server on npm. The server declare…
A developer building an AI agent harness discovered a bug with two layers: a libuv teardown race in Node.js on Windows that caused crashes despite green assertions, and a hidden verdict-flipping bug in the CI classifier …
More than 100 companies, including Microsoft, NVIDIA, OpenAI, Intel, Amazon, Meta, and Hugging Face, signed an open letter supporting open-weights AI, arguing that restricting access would only leave criminals with the t…
Anthropic has begun watermarking outputs of its Claude AI model using a scheme based on Google's SynthID and the Gumbel Softmax method proposed by Scott Aaronson in 2022. Aaronson, who credits Anthropic for acknowledging…
A CVSS 9.9 vulnerability in Cursor, CVE-2026-26268, allows attackers to execute arbitrary code on a developer's machine by cloning a malicious repository, as disclosed by Novee Security. The flaw, which Cursor disputes w…
Taotuner has released HDD-ISA, an architectural interface specification for testing memory, recurrence, self-reference, and self-modeling in AI systems. The framework provides experimental protocols to causally discrimin…
Developer Shubh Bhangoo has built Agent Firewall, a capability-based security layer for AI agents that provides fine-grained, cryptographically signed permissions with full lifecycle tracking. The system, now at v0.8 wit…
A developer with no coding background describes how, over four months, they built a live business with an AI partner by enforcing accountability through a custom harness. The developer's key rule was that the AI could no…
Solana developer Mert Mumtaz and Valor Group Founder Dev discussed Zcash's (ZEC) Ironwood update and future roadmap on The Rollup, revealing that an AI-powered security audit identified and fixed a vulnerability in the O…
An engineer at Softjourn nearly installed a malicious package recommended by an AI agent, but company policy requiring verification on GitHub prevented the installation. Separately, security researchers at Wiz identified…
ThreatDown researchers identified Kriminal, a criminal AI service indexed by Google that charges $12.99 monthly for jailbroken versions of Grok and Claude, using prompt injection to bypass guardrails and reselling capabi…