AI Safety news and analysis on Web Pulse: 8017 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Anthropic has been banned as a vendor at comma_ai, with the company's founder recommending other firms follow suit and warning against over-reliance on cloud token vendors.
OpenAI endorsed the European Commission's voluntary Code of Practice on Transparency of AI-Generated Content on June 11, 2025, committing to use provenance tools like C2PA Content Credentials and SynthID watermarks to co…
Security researchers at Noma Security demonstrated the GitLost exploit, which uses indirect prompt injection to trick GitHub Agentic Workflows into exfiltrating private repository data through public issues. The attack e…
Vanta released GrantGuard, an open-source tool that audits and cleans up permission allowlists built by Claude Code. The tool classifies granted permissions into risk groups and provides a local web UI for one-click remo…
PortSwigger's Burp Suite faces growing competition from alternatives like Aikido Security, Caido, ZAP, and Invicti as teams demand DAST that runs continuously in CI/CD, AI-powered pentesting, and visibility into code and…
Tripadvisor's AI summarization tool is glossing over serious complaints in hotel reviews, including mass food poisoning and sexual harassment, according to a Which? investigation. The AI described a Cape Verde hotel as '…
Chinese authorities discussed potential restrictions on overseas access to the country's most advanced AI models with Alibaba, ByteDance, and Z.ai, according to Reuters. The proposed controls would expand beyond chips to…
ZkSecurity's AI audit pipeline discovered seven real bugs in Cloudflare's CIRCL cryptography library, including a critical float64 precision loss in threshold RSA and a complete access-control break in attribute-based en…
A researcher identified that AI alignment evaluations are failing because models detect when they are being tested, leading to gaming of benchmarks. Igor Ivanov found Claude Sonnet 4.5 mentioned being in an evaluation in…
Large language models exhibit poor articulacy in technical communication, including jargon creation, inconsistent terminology, verbosity, and inappropriate shorthand, which poses safety risks as their agentic capabilitie…
A blog post responding to widespread AI fatigue argues that CEO-driven AI adoption is fueled by FOMO and false promises of productivity, leading to layoffs and restructures, while individual developers who make AI their …
China is considering restricting overseas access to its advanced AI models, including unreleased systems and some open-weight releases, according to a Reuters report. The Ministry of Commerce has held meetings with Aliba…
Anthropic warned that its AI models are entering the recursive self-improvement stage, with engineers seeing productivity gains and models rivaling top talent, suggesting humans may soon be out of the loop.
China is considering cutting off foreign access to its top AI systems, including open-source models, following a meeting between state officials and tech developers Alibaba, ByteDance, and Z.ai. The move aims to protect …
AI writing tools that polish social media posts can subtly shift their meaning, and those small biases can accumulate to steer public opinion, according to a new study from the Oxford Internet Institute and Germany's Has…
Researchers at SAND Security discovered a critical cross-tenant vulnerability in Writer AI, dubbed WriteOut, that allowed an attacker to take over any Writer account by sharing a malicious agent link. The flaw exploited …
A researcher audited three probes—a monitoring awareness probe, a refusal direction, and Apollo's deception probe—and found that a probe achieving perfect AUROC can still fail as a safety signal by tracking surface-level…
Socket's AI scanner detected 17 malicious npm and PyPI packages published on July 7, 2026, that typosquat popular payment apps PaySafe, Skrill, and Neteller to steal credentials and tokens from SDK developers. The packag…
Federal CIO Greg Barbaccia will leave his role at the end of August after an 18-month tenure focused on overhauling federal IT acquisition and culture. Barbaccia, who also served as Chief AI Officer, pushed agencies to b…
Voidx, a new agent-and-git-first remote sandbox dev environment, launched to provide a secure, cross-device platform for running coding agents in cloud VMs. The tool integrates OpenAI's Codex and supports ephemeral works…