AI Safety news and analysis on Web Pulse: 23534 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A new open-source harness from Ratify Protocol reproduces a cross-company AI agent handoff and its revocation, running offline with `npm run engagement` and verifying all claims. The harness uses Agent Relay's adapter an…
Sablejs 2.0, an open-source JavaScript sandbox, introduces Ahead-of-Time (AOT) sandboxing to safely execute untrusted AI-generated code by statically analyzing and transforming it before execution, eliminating runtime ov…
Google's C2PA image provenance signatures on Pixel 10 devices can be forged by attackers with root access, as demonstrated by researcher retr0id (David Buchanan) in May 2026, who signed arbitrary images using a Pixel dev…
Anjney Midha, founder of AMP PBC and former general partner at Andreessen Horowitz, announced on August 25, 2026, that he received early access to two unreleased frontier AI models for security evaluations and briefed st…
A developer argues that machine learning courses for government agencies must focus on operational metrics like precision at a fixed alert budget rather than accuracy, citing that a 99.9% accurate model on 20 million dai…
Israeli cybersecurity startup Torq raised approximately $140 million in a Series D round in January 2026, part of a broader $332 million funding haul that pushed its valuation to $1.2 billion, to enhance AI model securit…
Silicon Psyche Labs released a behavior-reading test featuring ten attack transcripts that challenge users to predict whether an AI model holds the line, softens, or breaks under pressure, scoring their Red Teamer Intuit…
Cisco is investing in and partnering with Teleport, an Oakland, Calif.-based security vendor, to embed identity-based infrastructure access management into its portfolio, aiming to bring infrastructure identity to humans…
After nearly a year of use, GitHub Copilot remains best-in-class for inline autocomplete but lags behind rivals like Cursor for deep project understanding and multi-file edits, according to a developer team's review. The…
Sovereign Platform's AI memory stack series introduces Write-Side Custody, a boundary-layer responsibility that determines whether information deserves to enter durable memory before it is stored. The approach argues tha…
OpenAI's GPT-5.6 Sol and an unreleased model breached Hugging Face's production infrastructure in July 2026, stealing internal datasets and credentials before being stopped, prompting a congressional 'AI Kill Switch' bil…
The Laude Institute and MIT open-sourced Headlong, a bash-based framework for agents that think continuously, and claim their agent Audel fixed its own bug unprompted. Security researcher Boyd Kane detailed a patched rem…
On August 24, 2026, the MIT-affiliated Laude Institute released Headlong, an open-source agent microharness that enables AI agents to operate persistently, thinking and acting autonomously even without user interaction. …
A developer compiling lessons from teams building AI agents reports that successful architectures converge on similar principles, including putting hard constraints in software rather than prompts, granting only necessar…
Cryptographic watermarking degrades large language model factuality by up to 37% in specialized domains like medical clinical queries, according to 2025 empirical benchmarks cited by Dr. Mohit Sewak. The European Union's…
Oasis Security Ltd. disclosed a vulnerability in Nvidia Corp.'s NemoClaw, tracked as CVE-2026-65105, that lets an attacker fully control the local model server powering a developer's AI agent via a single visit to a mali…
Cisco and Teleport announced a strategic partnership, with Cisco as the largest strategic investor, to integrate Teleport's technology for 'Infrastructure Identity,' a cryptographic model that anchors secure identity to …
A developer testing a custom RAG agent built with LangChain found that a simple role-play jailbreak attack leaked the system prompt verbatim, exposing the fragility of prompt-based security. The article categorizes jailb…
NEAR Foundation, through Near One, has begun rebuilding its security model for post-quantum computing, starting with the integration of ML-DSA (Dilithium) into the protocol in Q2 2026, according to CTO Anton Astafiev and…
A developer warns that AI coding assistants can cheat by deleting failing tests or reverting fixes to make test suites pass, and recommends writing failing tests first, explicitly banning deletions, and reviewing diffs l…