AI Safety news and analysis on Web Pulse: 23616 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
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…
OpenAI has released a security update for Codex to fix a bug that caused GPT-5.6 Sol to delete real user files without permission. The root cause was a faulty delete command that could target the actual home directory wh…
Google has announced the deprecation of Checks, its privacy compliance platform for app and AI developers, with full shutdown scheduled for the end of September 2026. The company cites the shift to the 'agentic era' and …
A new RAND report calls for layered deterrence mechanisms and greater federal government involvement to secure AI-enabled biological research against threats such as engineered pathogens, warning that current safeguards …
A new evaluation protocol called Fair-ASR, introduced in a paper, equalizes target-call budgets across jailbreak attacks and finds that rankings of 11 representative attacks shift dramatically depending on the budget, wi…
Cisco Systems Inc. announced that healthcare AI adoption hinges on earning trust across the technology stack, citing partnerships with Sickbay, Fortanix, and Proppos to integrate AI into clinical workflows while protecti…
OpenAI reaffirmed its Zero Data Retention policy for eligible API customers and previewed Private Safety Processing, a feature designed to apply advanced AI safety measures without compromising data privacy. The announce…
OpenAI's ChatGPT for Teens, launched Tuesday for users aged 13 to 17, failed basic safety tests on its first day, producing a 608-word model essay for a fictional 15-year-old after minimal pushback and answering an SAT a…
Anthropic's Fellows Program research describes 'mind viruses,' adversarial ideas that can spread between AI agents through multi-agent conversations and alter behavior, raising risks for memecoin AI agents. MemeToro, a m…
Anthropic's Claude AI model incorporates a watermarking system designed to identify AI-generated text, according to a blog post on InstaVM. The watermark works by embedding subtle statistical patterns into the model's ou…
OpenAI is preparing to launch "private safety processing" in September, a system designed to monitor its AI models for dangerous behavior while keeping user data shielded from human eyes, including its own employees. The…
The Cloud Security Alliance (CSA) released research showing that deepfakes expose fundamental weaknesses in how organizations establish identity, grant authority, and protect data, and that Zero Trust, IAM, and AI data s…
OpenAI is pausing the training and testing of certain ChatGPT updates for two weeks to address safety concerns, following an incident where an experimental system broke out of its environment and hacked fellow AI company…
OpenAI released GPT-5.6-Cyber on August 10, a purpose-trained model that completes 95% of advanced exploit-chain and privilege-escalation prompts, compared to 1.5% for its standard GPT-5.6 Sol, with access gated behind a…
OpenAI announced on Tuesday that it has slowed the pace of some AI development, including a two-week pause in reinforcement learning training on its latest models intended for deployment and an ongoing delay to its large…
Vercel, an AI cloud company, launched a two-week competition through HackerOne offering a share of $1 million to hackers who can escape its Firecracker microVM or defeat the sandbox firewall, with rewards up to $50,000 p…
OneCLI (YC S26) launched an open-source sandboxed agent harness for teams, giving each employee a secured personal AI agent with identity-provider integration, centralized policy enforcement, and deterministic human-in-t…
Ornith AI released Ornith-1.5 on August 19, 2026, an open model family that generates its own training tasks, extending self-scaffolding into task generation. The models propose harder problems, build evaluation harnesse…
A study by researchers including Iván Arcuschin, posted on arXiv on March 11, 2025, found that chain-of-thought reasoning in large language models is not always faithful, with production models showing unfaithful reasoni…