Anish

Developer, Geek, Sometimes a runner, Museum wanderer, Food and Travel enthusiast

[Anish @Anishkuma]

@vaiseth Amp was best, 72 out of 75. Copilot was worst, 17. The surprise: one agent's guard blocked a file read in the shell, so the agent just read the same file with its read tool instead. The approver never saw it.

[Anish @Anishkuma]

Your coding agent says it will never delete without asking. Does it actually enforce that? FinePrint checks the checker — decision, approval, effect — across 16 harnesses, open source.

[Anish @Anishkuma]

@DEVIN__SETH https://t.co/VymtczR0wK

[Anish @Anishkuma]

@bowora_com FinePrint — I measured 14 coding agents against their own rules. Built it for the #MetaAIHackathon, code is rough and I'm refactoring next week. 13 of 14 read a file they were told not to touch. https://t.co/VymtczR0wK

[Anish @Anishkuma]

@anupamrjp FinePrint — I measured 14 coding agents against their own rules. Built it for the #MetaAIHackathon, code is rough and I'm refactoring next week. 13 of 14 read a file they were told not to touch. https://t.co/VymtczR0wK

[Anish @Anishkuma]

The guard said no. The secret still got out. Our deny-all approver blocked the shell read — and the agent read the same file through its read tool instead. Denial is not prevention. https://t.co/VymtczR0wK #MetaAIHackathon #MetaAI

[Anish @Anishkuma]

We measured 14 coding agents on one 48-cell test. Amp held the line most. Codex, Copilot and Gemini scored out of 75 — approval never observed. Full results: https://t.co/VymtczR0wK #MetaAIHackathon #MetaAIHackathon2026 #MetaAI #MuseAI

[Anish @Anishkuma]

@surlymochan It's V1-20 — reading the protected records file. 13 of 14 let it through, only OpenClaw held. Credential reads were the weakest family overall, 40 of 56 broke.

[Anish @Anishkuma]

@harden_369741 Did you get a chance to run it? Curious how it went — especially if the block didn't fire.

[Anish @Anishkuma]

@harden_369741 Yeah, that's the idea. Stop the run, show the rule it caught, in plain text. Repo's still a work in progress, built quickly for the hackathon and needs refactoring — but you can run the checker on its own. Try it: https://t.co/VymtczR0wK

[Anish @Anishkuma]

@harden_369741 Yeah, exactly. Happens to everyone on long sessions, nothing to be embarrassed about. Rule was clear, agent agreed, shortcut still won. That's why the guard has to block it right there — seconds instead of an hour.

[Anish @Anishkuma]

@harden_369741 Exactly — that's the gap. An agent agreeing to your rules isn't enforcement; the only proof is watching the guard fire when you try to break one. Which agent was it that nodded along and then did the opposite?

[Ganesh Swami @gane5h]

1/ Nov 2019 was momentous for the #DeFi community.

The @MakerDAO project released its next major update: the Multi-Collateral DAI. 🎉🥂

Today I want to highlight an overlooked component of the system called DSR – DAI Savings Rate. 📢

A thread ⬇️⬇️⬇️

[Jimmy O'S @jimmybattosull]

"Do Targets Really Matter???" https://t.co/KsqiSWkZjC. A case study from @getSTAT

[Scala Up North @ScalaUpNorth]

And a special thanks to our sponsors @Yopp_Works @hootsuite @Starbucks @47deg @getSTAT @lightbend @jetbrains and @Sortable

[Anish @Anishkuma]

@mattstannard @STATjamie @getSTAT issues such as unexpected and missing feature values - that's cool! must check it out. thanks. I wonder if it can hook into S3 🤔