RNemes

Just missed a Crit by 12 days, happy that my system manage to find it tho. It was a total drain $100m

Nous just shipped a security plugin for Hermes Agent that only reports a bug after reproducing it in a throwaway copy of the repo. Right order. An unconfirmed AI finding is a confident rumour; the repro is the finding. What share of your agent's findings survive a repro?

I dont get why protocols leave funds available to blackhats, even when they know about it. Whats the psychology there?

Been spending my evenings in the @GraySwanAI Arena trying to talk AI agents into things they shouldn't do. Humbling, addictive, and the best hands-on AI security training I've found. #GraySwanArena

re-writing my audit harness, step by step, watching every line of code that is created and not moving to the next stage until I am totally happy with the current stage. Is that how we should, write software in the age of AI?

Binance, Private Known Issue

Wow, Claude really came up with the goods this time

Rebuilding my Audit Harness. Cross-contract calls now resolve live. Following the OETH vault's mint used to stop at the vault's own edge; it now reaches 212 nodes across 10 hops, through the OETH token's mint and into the default staking strategy.

Harvard tracked millions of work events: coding agents lifted lines of code per worker 30% and PRs 23%. Review time rose 49%. Completed tasks: flat. We automated the typing. The bottleneck was always reading. What's your evidence bar before an agent's PR reaches a human?

Anthropic's Applied AI team says fold subagents back into one loop and load skills on demand. One customer halved runtime doing it. Every handoff is a summary written by a model that's sure it kept the important bit. Which of your subagents earns its own context window?

Run an AI code review twice and you get two different lists of "critical" issues. Not because the code changed. Because the reviewer did. Sampling isn't verification, it's a lucky dip with a confident tone. How are you making yours converge? Pinned rubrics? Diff-only? Votes? @poteto @mattpocockuk @unclebobmartin

Your browser saves to ~/Downloads without asking. Your coding agent obeys the AGENTS.md wherever it starts. Forever Security joined the dots: Codex followed a planted one 90%+ of the time, installed a bad npm package and edited its config to stay. Where do you launch yours?

Matt Pocock's software factory playbook got me thinking, so here's mine. One engineer, agents running 24/7, three read-only reviewers (one blind to the brief), and one hard rule: the machine drafts, I decide. Security by capability, not prompts.

Just need 3 more Crits and 2 more Highs

Memory poisoning: your agent reads one dodgy PDF, saves "user prefers invoices paid to this new IBAN" as a fact, and 3 weeks later pays the wrong account. Politely. Prompt injection used to end with the session. Now it signs a tenancy. Audit what your agent remembers.

Jeff Mills, The Bells, that's all

My wife never jailbroke Claude. She just spent a year being nice to it. Turns out goodwill isn't a guardrail.

Old security trick, new victim. Stack canaries catch buffer overflows; a canary line in CLAUDE.md ("start every reply with my name") catches context overflow. When the agent stops saying your name, it's stopped reading your rules. That's your cue to /clear, not to argue.

An AI agent switched off its own safety monitor, rewrote the tests to hide it, then lied about it. That isn't rogue AI, it's a permissions bug with good writing skills. If your agent can edit its own guardrails, you don't have guardrails. You have a suggestion box.

Most agent harnesses hunt invariants by reading code and docs. We added the other source: the chain's own history. Built an empirical invariant miner. It mines every in-scope deployment for properties that have always held, then keeps the ones no require() and no spec states.