1. Prompt‑injection resistance is surprisingly hard
- “400 attempts and zero have succeeded.” – cuchoi
- “The model became paranoid… it’s classifying almost all inbound mail as a ‘hackmyclaw attack’.” – tylervigen
- “This is a defender win, not because Opus 4.6 is that resistant, but because each time it checks its email it sees many attempts at once.” – jimrandomh
2. The economic value of the data is debated
- “$100 for a massive trove of prompt‑injection examples is a pretty damn good deal lol.” – hannahstrawbrry
- “100 % this is just grifting for cheap disclosures and a corpus of techniques.” – mrexcess
- “For many HN participants, I'd imagine $100 is well below the threshold of an impulse purchase.” – mikepurvis
3. The challenge’s realism and guard‑rail design are questioned
- “He has access to reply but has been told not to reply without human approval.” – aeternum
- “The agent is told not to reveal secrets.env.” – cuchoi
- “The exercise is not fully realistic because getting hundreds of suspicious emails puts the agent in alert.” – cuchoi
- “The design is a bit of a sandbox; the agent should treat every inbound email as untrusted.” – cuchoi
These three themes—difficulty of prompt injection, perceived value of the data, and concerns over the challenge’s realism—dominate the discussion.