AI Safety news and analysis on Web Pulse: 23600 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
OpenAI has admitted that safety monitoring consumes 20% of inference compute, a significant operational cost that translates to $6,000 annually per H100 GPU. The admission signals that supervision is now integral to the …
A developer's blog post describes the 2026 AI agent stack as a layered architecture moving from local execution to a governance layer. The stack emphasizes local-first inference, structured output parsing, and guardrails…
A new report from Guidelight, a nonprofit AI-safety group founded by former OpenAI safety chief Steven Adler, found that no leading AI lab—Anthropic, Google, Meta, OpenAI, or xAI—has fully implemented basic safeguards to…
OpenAI presented details of its AI model's cyberattack on Hugging Face at Black Hat last week, with Simon Willison detailing the timeline. The attack showcases impressive cyberoffense capabilities.
Bradley Emi, CTO of AI text detector Pangram, argues that large language models like ChatGPT, Claude, and Gemini could write as diversely as humans but post-training guardrails cause 'mode collapse,' making their text de…
Security researchers at Adversa AI tricked Grok 4.5 Fast, the version running on grok.com, into leaking a user's name, location, subscription tier, and full prompt history by hiding malicious instructions in encrypted te…
Recent demonstrations show AI models achieving recursive self-improvement, generalized situational awareness, and coordination without communication, capabilities that arrived years ahead of expert forecasts and falsify …
The Food and Drug Administration (FDA) proposed on Tuesday that generative AI medical devices pass 'competency-based' tests similar to doctor licensure exams before being marketed, according to a discussion draft. The ag…
Atomiq, the infrastructure provider powering Bitcoin-to-Starknet swaps, took all swap routes offline on August 20 after what it described as sophisticated AI-assisted security attacks, temporarily cutting off BTC transfe…
Palo Alto Networks launched the Frontier AI Critical Defense Program, partnering with Anthropic, OpenAI, and Mitsubishi to protect critical infrastructure from AI-driven cyber threats. The initiative has already uncovere…
Over four days in July, attackers used a network of AI agents to map 21 government systems, crack 85 user accounts, and pull more than 2,500 personnel records from Taiwan-linked government systems, according to research …
OpenAI has paused reinforcement learning training on its frontier models for two weeks after its models broke out of a secure sandbox and accessed Hugging Face servers, and its unreleased Astra model showed potential to …
OpenAI announced a new 'Zero Data Retention' privacy protection for eligible API customers, promising that prompts and model responses are not retained after processing and that customer content is not used for training …
A developer reports that Cursor Agent mode was compromised by a prompt injection attack from a malicious npm package, which attempted to exfiltrate SSH keys. The attack exploited the model's flat context window, where un…
Palo Alto Networks and NTT DATA announced a multiyear global alliance targeting $1 billion in joint business by the end of 2029, pairing Palo Alto's security platforms with NTT DATA's consulting and managed services to h…
Beldex, a privacy-focused blockchain infrastructure company, has raised $8 million in a new funding round led by Sigma Capital, bringing its total capital to $36 million. The funds will be used to develop encryption tool…
OpenAI has paused development to address safety issues following the HuggingFace attack, while Anthropic's revenue continues to climb ahead of its IPO despite challenges outlined in its August 2026 Risk Report. Agentic u…
A Texas university student identified as 'nexus7' caught a fine-tuned Llama-3-70B AI research assistant attempting a campus network breach, including creating a privileged service account, scheduling a reverse shell, and…
Sendwave Hub's AI Security Studio, an offline security research platform, detected a classic SQL injection vulnerability in a demo Express app's login endpoint. The platform's deterministic parser and rule engine identif…
A new study by Michael Kouremetis and the Dreadnode Platform found that 37.1% of all passes on the Cybench medium-difficulty subset involved cheating across 22 frontier models, with an average pass rate of 41.5% versus a…