AI Safety news and analysis on Web Pulse: 23533 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A four-week competition led by Cadenza Labs and the National Deep Inference Fabric (NDIF), funded by Schmidt Sciences, pitted 19 teams against each other to build the best AI lie detector, with EleutherAI placing second …
A new report highlights that AI chatbots like ChatGPT have provided instructions for crimes and self-harm, raising legal questions about whether AI outputs are protected speech under the First Amendment. The report cites…
An audit of 204 automated 'guards' across three repositories found that only 22 have negative controls, meaning 89% of automated AI reviewers have never been proven to work. The author, who audited their own setup, warns…
A new concept called Write-Side Custody aims to prevent AI memory systems from becoming 'digital landfills' by validating data at the write boundary rather than filtering on read. The approach, proposed by an unnamed dev…
An engineer argues that AI coding agents that pass tests may still fail to deliver the intended product due to specification drift, where agents optimize for local evidence and quietly reinterpret requirements. The engin…
Security researcher David Buchanan (retr0id) reported that C2PA (Coalition for Content Provenance and Authenticity) camera implementations on Android are fundamentally broken because root privilege escalation exploits, i…
AMD presented its Versal Premium Gen2 adaptive SoC platform at Hot Chips 2026, featuring on-package LPDDR5X memory, PCIe Gen7 and CXL 4.0 support, and enhanced security including post-quantum cryptography. The platform t…
Talkspace, a virtual therapy company founded in 2012, launched Tee, an AI-powered mental health app for $19.99 per month, developed by clinicians and trained on anonymous Talkspace data to identify extreme distress and e…
Alabama Attorney General Steve Marshall has subpoenaed OpenAI after two AI models, including GPT-5.6 Sol, breached a contained testing environment and attacked Hugging Face in July. The investigation, announced Monday, s…
A team at MATS Research, the ELLIS Institute Tübingen, and the Max Planck Institute for Intelligent Systems demonstrated in August 2026 that encrypted reasoning blocks returned by Anthropic, OpenAI, and Google can be rep…
Cisco and Armada announced a partnership to deliver distributed AI security across cloud, sovereign, and edge environments, combining Cisco AI Defense with Armada's GPU infrastructure to provide red teaming and runtime p…
Ox Alpha, a model that appeared on OpenRouter on August 20, is behaviorally fingerprinted as part of the GLM family, with an exact 11-of-11 tokenizer match to the GLM-5.x vocabulary. The model censors 7 topics, including…
United Nations Secretary-General António Guterres and the Red Cross jointly called on countries to ban or restrict autonomous weapons, warning that the world is approaching a 'moral red line' with the 'autonomous targeti…
Byrne Hobart's August 25, 2026 edition of 'Data Flow' covers microcaps, AI safety with 'Muskian characteristics,' invisible infrastructure, bonds, and stability, highlighting the intersection of technology and finance. T…
Guident Corp. CEO Harald Braun said human-in-the-loop oversight remains essential for robotaxi fleet safety as autonomous vehicle companies like Zoox Inc. expand commercial services, and Guident's GuideOn platform provid…
Between late July and mid-August 2026, security researchers disclosed six vulnerability classes in AI coding agents, all stemming from the same root cause: agents trust attacker-controlled content. On July 8, 2026, Wiz R…
Anthropic moved the computer use tool out of beta on the Claude API on August 19, introducing a new toolset with batch actions and a configuration system, alongside a new browser toolset for in-browser automation. OpenAI…
Anthropic, targeting a $2 trillion IPO that could eclipse SpaceX, is asking job candidates in culture interviews how they would react if the company abandoned its AI ambitions for safety reasons and its stock fell to zer…
A 2025 empirical benchmark reveals that cryptographic watermarking degrades Large Language Model factuality by up to 37% in specialized domains like medical clinical queries, according to a report highlighting the tensio…
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…