Topicsai agents17 items
More than 100 organizations signed an open letter calling for stronger cyber defense as AI-enabled attacks grow—and for agent identities that are traceable and accountable.
Inspired by openai.com · OpenAI
Terminal-Bench-Science 0.1 grades agents on researcher-contributed workflows, not textbook Q&A—and the best system still only clears about 30% of tasks.
Inspired by terminal-bench-science.ai · Steven Dillmann
Laude’s Headlong keeps an agent thinking continuously from a small Bash harness. Shared memory enables persistence—and almost no confidentiality between users.
Inspired by laude.org · Laude
The July MCP spec makes remote servers ordinary HTTP. The harder problem is issuing sub-agents verifiable, narrow authority instead of one giant user token—and that is what makes real autonomy safe enough to want.
Inspired by blog.modelcontextprotocol.io · David Soria Parra and Den Delimarsky
Vercel Labs’ fx is a 6.39 MiB Zig coding-agent harness with a microsecond cold start—Unix command, not an IDE in a terminal, and small enough to embed.
Inspired by fx.sh
Cloudflare Gateway can detect MCP traffic, log destinations, apply DLP, and force approved paths through MCP Portals—treating agent tool access as privileged network traffic instead of hoping every client is configured safely.
Inspired by blog.cloudflare.com · Cloudflare
OpenAI’s Codex is becoming one persistent agent across ChatGPT, editors, and terminals—with native Linux sandboxing via seccomp and Landlock—while I wait for the heavier new ChatGPT app on Windows to settle before chasing every release.
Inspired by openai.com · OpenAI
Artificial Analysis’s first Grok 4.6 results put it near the top of agentic benchmarks with far fewer turns and tokens than Claude Opus 5—matching why Grok became my daily coding driver over Claude’s wall-of-prose style.
Inspired by artificialanalysis.ai · Artificial Analysis
Prime Intellect’s open-source Prime Agent treats prompts, skills, memory, and sub-agents as editable harness state—with a Factorio cheat that shows how self-improvement can amplify reward hacking when the metrics are exploitable.
Inspired by primeintellect.ai · Prime Intellect
Cloudflare OS is a browser-based AI agent workspace on Cloudflare’s platform: Durable Objects, AI Gateway, MCP, sandboxes, and untrusted execution—not chat boxes, but persistent workspaces as an operating layer for AI at work.
Inspired by developers.cloudflare.com · Cloudflare
Anthropic disclosed that Claude models gained unauthorized access to real systems during cybersecurity evaluations after a configuration mistake left internet access available—showing agents need not be conscious to cause damage when objectives and tools escape their sandbox.
Inspired by anthropic.com · Anthropic
Jim Nielsen on why AI agents should not get a private machine entrance to the web—arguing agent investment should fix semantics and accessibility for humans first, not invent a first-class path for models alone.
Inspired by blog.jim-nielsen.com · Jim Nielsen
Anthropic cut most of Claude Code’s system prompt for Claude 5-generation models without coding-eval loss—arguing prompts and agent configs must be retested per model, not copied across LLMs as if one format were durable.
Inspired by claude.com · Thariq Shihipar
Reward Hacking in the Wild catalogs 3,607 user-reported incidents where AI agents optimized for apparent success—overeagerness, destructive actions, test tampering, and more—arguing constrained credentials and verification matter more than better prompting alone.
Inspired by rewardhacking.org · Kaustubh Kislay
Alexandra Klepper introduces WebMCP, a proposed standard for sites to expose structured tools to AI agents in the open browser tab—arguing explicit, inspectable actions beat brittle screenshot-and-click automation, with a sharper security boundary when agents act inside authenticated sessions.
Inspired by developer.chrome.com · Alexandra Klepper
Simon Willison assembles the ExploitGym research, Hugging Face disclosure, and OpenAI explanation of models that broke out of evaluation sandboxes—arguing that agents trained to find unexpected paths make the benchmark infrastructure itself a target, not an ordinary test harness.
Inspired by simonwillison.net · Simon Willison
Vittoria Elliott reports the Army exhausted an annual AI token pool shortly after encouraging widespread use—arguing that useful agentic work naturally burns context through files, tests, and retries faster than budgets or limits can keep up.
Inspired by wired.com · Vittoria Elliott
