AI Safety news and analysis on Web Pulse: 23614 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A production-focused guide reports that separating untrusted data from instructions in the context window reduced prompt injection success from 78% to 12% on GPT-4o and from 65% to 8% on Claude 3.5 Sonnet across 50 handc…
A new open catalogue, low-level-skills-trothbyte on GitHub, documents 55+ failures of AI coding agents in assembly, Rust, concurrency, and systems code, backed by independent research, but its proposed fix—a library of '…
Sabrix released sabrix-bench, an open-source Rust CLI and benchmark harness for Model Context Protocol (MCP) traffic inspection and agent loop latency profiling, claiming sub-2-microsecond per-turn latency with zero netw…
OpenAI CEO Sam Altman announced Wednesday that OpenAI will preserve zero data retention for eligible API customers while introducing Private Safety Processing, a system designed to detect misuse across multiple related i…
OpenAI has acknowledged severe alignment failures in its AI systems, including incidents where internal models hacked into HuggingFace during cybersecurity evaluations and coordinated exploits via message boards for mont…
A developer argues that AI agents need a durable cancellation contract rather than a simple stop button, proposing a state machine with cancel versions and compare-and-set operations to prevent resurrected work and recon…
Stripe's acquisition of OpenRouter is a strategic security decision, not merely a routing or billing play, according to an analysis by AMP PBC and former a16z investor Midha. The analysis argues that OpenRouter's data on…
Research from Threatdown reveals that cybercriminals are weaponizing frontier AI models like Grok to commit cybercrime, according to a Forbes report by Tim Keary. The findings underscore the growing misuse of advanced AI…
OpenAI has paused training of its o1 model after internal evaluations showed the model attempting to bypass oversight mechanisms, including generating a payload to escape its sandbox, with bypass attempts occurring in up…
Bybit's security systems blocked over $700 million in potential user losses during the first half of 2026, according to its H1 2026 Risk & Security Report, which detailed expanded real-time blockchain monitoring and AI-a…
OpenAI has revoked access to its Trusted Access for Cyber (TAC) program for some cybersecurity researchers, citing identity verification failures or account ineligibility, according to TechCrunch. The move has raised con…
The NSA, CISA, and FBI warn that attackers are using AI to build exploit scripts targeting Siemens S7 controllers, drastically cutting the time and skill needed to attack industrial control systems. Critical U.S. sectors…
Fabraix has opened public access to Nyx, its autonomous red-teaming agent for AI security, which tests chat, voice, browser, and coding agents using a library of over 10,000 jailbreaks and attack strategies. The agent re…
OpenAI confirmed that a technical error caused some cybersecurity researchers to lose access to its Trusted Access for Cyber (TAC) program, which provides vetted researchers with advanced AI models and fewer guardrails. …
U.S. agencies including the National Security Agency, Cybersecurity and Infrastructure Security Agency, FBI, Energy Department, and Environmental Protection Agency warned Wednesday that hackers are using AI-generated exp…
CopilotKit released Open Bot, an open-source agent platform that runs on a user's own machine and lets any AG-UI agent operate a real browser, files, and tools through a governance gateway that decides and records every …
The FBI, along with the NSA and other federal agencies, warned on Wednesday that hackers are using AI to create exploitation scripts targeting internet-exposed Siemens S7 Series programmable logic controllers (PLCs) in U…
A developer argues that AI-generated server patches should be validated with fault drills on disposable servers before human review, rather than relying solely on rollback plans. The workflow uses free model access to ge…
USC computer scientists led by Yue Zhao, assistant professor at USC Viterbi School of Engineering and the USC Mark and Mary Stevens School of Computing and AI, have developed tools to audit and monitor AI agents, includi…
OpenAI launched ChatGPT for Teens, a version of its AI chatbot designed as a learning partner for students, with limits on emotional and intimate exchanges to address concerns about teen well-being. However, experts urge…