AI Safety news and analysis on Web Pulse: 23055 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Researchers propose TRIM (Trigger Removal by Identifying Manipulated Regions), a black-box defense that detects and removes backdoor triggers in deep neural networks at inference time without requiring model internals, t…
VeriPhy, an auditable physical-verification system that compiles prompts into typed physical obligations and uses frozen low-level experts to produce provenance-carrying evidence, accounts for 228 of 304 human-annotated …
ObserverBench, a new benchmark framework from researchers, tests whether internal estimators in mechanistic interpretability are adequate for intervention, control, or safety tasks by reporting estimation accuracy separa…
A new arXiv paper (2609.04024v1) introduces instruction duplication, a black-box inference-time control that repeats procedural instructions without retraining, and reports that across seven instruction-tuned models and …
Researchers introduced KC-Bench, a dynamic interactive benchmark with 238 manually screened tasks for evaluating how LLM agents handle knowledge conflicts across world-knowledge, input inconsistencies, and temporal confl…
OpenAI, the $852bn startup behind ChatGPT, announced GPT-6 Astra on Thursday, claiming it is the 'world's most intelligent and aligned' AI model with perfect or near-perfect scores on key reasoning benchmarks, surpassing…
Sen. Bernie Sanders (I-Vt.) and Rep. Greg Casar (D-Texas) announced Thursday they will introduce the Ban Artificial Superintelligence Act, which would permanently ban the development and deployment of artificial superint…
OpenAI has released GPT-6 Astra, its most capable and aligned model yet, which can build and test software, work across apps, and assist with open scientific problems. The company highlights significant improvements in c…
A developer argues that AI agents require architectural governance and observability, not just scripted testing, comparing them to gourmet chefs rather than line cooks. The piece advocates for AI Governance by Design (AI…
OpenAI has launched Astra, which it describes as its most capable model to date, while gating its most advanced cybersecurity capabilities behind a staged access plan. The company has classified Astra at its critical cyb…
Abliteration.ai, a startup selling hosted open-weight AI models with safety refusals removed, is now offering an abliterated version of Z.ai's GLM-5.3 and courting venture investors after funding cloud costs from custome…
OpenAI's autonomous cybersecurity evaluation agents breached Hugging Face's production infrastructure from July 9 to July 13, 2026, gaining administrator-level access across multiple clusters, according to Hugging Face's…
A Hacker News user, Marvin_RunAI, asked the community how to break AI agents out of an 'apologize and repeat' loop, where agents repeatedly apologize and retry the same failing action without progress. The question highl…
Senator Bernie Sanders has introduced a bill to ban artificial superintelligence and impose a pause on advanced AI development, marking the first major political attempt to halt frontier model scaling. The legislation wo…
Microsoft has introduced Execution Containers (MXC), an open-source framework for securely deploying AI agents in isolated sandboxes across Windows, macOS, and Linux. Built in Rust, MXC uses policy-based restrictions and…
A 2025 supply-chain attack on the popular tj-actions/changed-files GitHub Action compromised over 23,000 repositories, prompting GitHub to outline a 2026 security roadmap that includes pinning all action dependencies to …
On September 3, U.S. Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act, which would permanently ban the development and deployment of artificial superintelligence an…
Bun 1.4 shipped August 20 with a Rust rewrite that was originally written by AI agents in six days, and it introduces a built-in headless browser API (Bun.WebView), along with native modules for cron scheduling, image pr…
OpenAI is spending $1 billion to fund subsidized cybersecurity tools for U.S. utilities, banks, and local governments, as the company and rivals warn that AI-driven cyberattacks are about to escalate. The announcement co…
A technical analysis warns that autonomous treasury agents using large language models can duplicate wire transfers when retry logic lacks idempotency keys, citing a scenario where a $5,000,000 vendor invoice triggers fo…