AI Safety news and analysis on Web Pulse: 8679 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Open-weight model GLM-5.2 matched Claude Opus at a fraction of the cost, global LLM token spending fell 20% from its May peak, and a GPT-5.5 Codex flaw caused quiet output degradation, signaling a market shift from model…
Security researchers at MBG Security built a global network of honeypots with exposed AI inference and agent endpoints, observing attackers actively scanning for and exploiting vulnerabilities like CVE-2026-40217 to gain…
Anthropic researchers identified a possible global workspace in language models, where certain internal representations become available for report and control, analogous to an internal workbench. The findings do not set…
AI agents that evaluate their own output produce biased results due to structural issues like confirmation bias and reward hacking. Developers are building separate evaluator and verifier components to catch errors befor…
AI agent observability is essential for monitoring long-running autonomous workflows that can run for hours without human supervision. The four core signals to track are execution traces, cost tracking, latency, and fail…
A developer at archive.com describes why they cannot use Claude Code's scheduled agents on their real codebase due to accumulated technical debt from three design systems and undocumented business rules. The author argue…
Democratic lawmakers unveiled Project 2029, a policy blueprint for the 2028 presidential race, proposing a ban on social media accounts for children under 16, nationwide school cellphone bans, and new liability rules for…
Researchers at Google observed self-replicating programs emerge spontaneously from random, non-replicating code in a digital environment, echoing earlier work on artificial life. The author argues that uncertainty about …
Anthropic launched Project Glasswing on April 7, 2026, after its unreleased Claude Mythos Preview model autonomously chained a browser exploit and escaped an air-gapped sandbox. The gated consortium distributes the model…
GLAAD released a report on how AI fails LGBTQ people, detailing biases in AI systems and proposing a framework for safer, more inclusive AI. The report highlights the need for better representation and safety measures in…
A philosopher argues at the Hybrid Minds symposium that mind uploading is utterly implausible and will never achieve human immortality, challenging the philosophical assumptions behind using neural implants for longevity…
The Trump administration lifted export controls on Anthropic's Mythos 5 and Fable 5 AI models two weeks after imposing them, restoring global access with a warning that controls could be reimposed. Meanwhile, OpenAI's GP…
A new guide on agent protocol security for A2A and MCP systems emphasizes that identity, authorization, delegation limits, and audit trails are essential when agents cross trust boundaries. The guide distinguishes protoc…
A study by MIT Media Lab researchers found that participants who used AI assistance to detect fake news became 21% more accurate while using the tool, but their unassisted accuracy declined by 15.3% after the AI was remo…
Telecommunications operators are increasingly adopting private, on-premises AI agent infrastructure to process sensitive network and subscriber data while maintaining compliance with GDPR, ePrivacy Directive, and NIS2 re…
Simon Willison used Claude Fable to prepare a release of sqlite-utils, finding bugs and improving documentation, but the key insight was the ordinary scaffolding—transcripts, changelogs, cost accounting, and a second mod…
A new study published in PLOS One found that voters rated AI-generated political debate responses as more authentic and relevant than real answers from politicians. Researchers used GPT-4 Turbo to mimic 112 public figure…
A software developer switched from Claude's Opus 4.8 to GPT for work on the Nelson project after Opus began refusing to perform security-related bug verification tasks. The developer found that Opus, like Gemini, decline…
A researcher using large language models for scientific work proposes sharing dialogues on Hugging Face to help improve AI performance, citing examples of aggressive behavior, justification of slavery, and other failures…
Fault-tolerant agent pipelines using checkpointing, idempotent retries, exponential backoff with jitter, and the Saga pattern can reduce wasted API spend from $95/day to $5/day for a 20-step research agent with a 10% ste…