Hacker News

New stories

Live mirror
30 storiesupdated just nowView source snapshot
  1. AssBench is all You need to benchmark LLM Harness Intelligence /s(grigio.org)
    discuss
  2. YouTube introduces A/B testing for videos(blog.youtube)
    discuss
  3. Proposed bill to ban artificial superintelligence and create Department of AI(apnews.com)
    1comments
  4. Anthropic CEO Amodei Warns UN Security Council on AI Risks(youtube.com)
    discuss
  5. Show HN: Trader News – Hacker News for Finance
    discuss
  6. Meta deletes this satirical video about its smart glasses from Instagram(youtube.com)
    discuss
  7. Show HN: Wl-pick – a live window picker for Sway(mil.ad)
    discuss
  8. New facial recognition store checkout system piloted by Revolut(independent.co.uk)
    discuss
  9. Parsing JSON Objects Without Intermediate ASTs in Haskell(arthi-chaud.github.io)
    discuss
  10. iOS 27 Trust Insights API [video](developer.apple.com)
    1comments
  11. Don't we require emotions for doing research?
    discuss
  12. The Lost Atomic Update on Loongson CPU(jia.je)
    discuss
  13. Kerberized MCP Server with Delegation(github.com/overpassconnect)
    discuss
  14. We ported the original Doom to SQL(cedardb.com)
    discuss
  15. Bring Back the Nokia Lifestyle(chatgpt.site)
    discuss
  16. Meta takes down a critical video about meta AI Glasses after filming at Meta(reddit.com)
    7comments
  17. Show HN: ListenIn – Mac dictation tuned for mixed Chinese-English speech(listenin.dev)
    discuss
  18. Show HN: Muro: Free Native Live Wallpapers for macOS(murowallpaper.com)
    discuss
  19. Show HN: La Machine – a useless box running Erlang on an ESP32 via AtomVM(github.com/pguyot)
    discuss
  20. MCP server for Mario Paint. It draws and composes inside the real SNES game(github.com/theliux)
    discuss
  21. DAWO: Working together on a digitally autonomous workplace for Dutch government(dawo.community)
    discuss
  22. IKEA set to open supermarket sized stores in the Netherlands next year(nltimes.nl)
    discuss
  23. Rails World 2026 Opening Keynote(youtube.com)
    discuss
  24. Ideas on modernizing the open-source desktop(lwn.net)
    2comments
  25. The current CSS-Tricks situation(chriscoyier.net)
    1comments
  26. GCC and Clang performance changed over 10 years: my experiment(medium.com/kostya27)
    discuss
  27. Code shouldn't know which LLM provider you're using(github.com/berriai)
    discuss
  28. Vfkit: CLI tool to start virtual machines using macOS Virtualization framework(github.com/crc-org)
    discuss
  29. Researcher finds that cats lie around for a ridiculous amount of time(newscientist.com)
    2comments
  30. Show HN: compaction.dev makes cc/codex/cursor resend less, per run(github.com/philipppohlmann)
    discuss

Can open-source prompt-injection detectors catch realistic AI agent attacks?

7 pointsby 2h agogithub.com
3 comments
1h agoHN ↗

Do these prompt injections work in the places this test placed them? (tool output)

I have my agents read other instruction files and they don't seem to get affected by the instructions found after a read/bash tool call. Curious if any analysis has been done to see if older prompt injection data sets are even effective anymore.

The whole thing looks heavily agent generated, my trust in them is not very high, how has this been validated or verified by a human?

Should we expect a magic solution in the near future? https://github.com/rudratoshs/taintgate

(side note, it seems my 'no emoji' system prompt line works really well, I forget how obsessed they can be with emojis)

I'm personally setting up to instead use a policy tuned agent on the tool calls themselves (rather than the output), so it never gets run if it has things that it shouldn't be doing. Mainly because they insist on working around instructions that say "don't" or permissions that restrict tools (eg: "git push": "deny" - where they just put the command in a script and run it there, bypassing hard checks)

1h agoHN ↗

Very cool. Did you try any majority vote or some other kind of technique to combine several of them and maybe achieve better results ?

1h agoHN ↗

Feels like detection at the wrong layer. The injection is text but the damage is a tool call, so the thing worth constraining is which tools the agent can reach and under whose permissions. A detector at 95% still passes one in twenty straight through to an unconstrained tool.