OpenCode under cyberattack: defense models refused to analyze exploits mid-incident; Daybreak access restored

3 people · 6 posts · first post Sun 15:13 UTC · latest 32 min ago

During a cyberattack on OpenCode, frontier models used for defense reportedly began refusing to help analyze the attacking agents, per Andrew Curran citing Dax; access was restored via Daybreak Blue. Dean Ball called it a modest, quickly resolved hindrance. David Sacks cited it as evidence against centralized SI-safety approaches in cyber.

All posts · 6, most important first

1. @andrewcurran_ commentary · Sun 15:13 UTC
   OpenCode is currently under a sustained cyberattack, and mid-incident they ran into the same problem Hugging Face did: the frontier models they were using for defense began refusing to help fight attacking agents because analyzing the attacks started triggering safety guardrails.
   
   yes that was a bit frustrating mid incident when trying to investigate. they did quite a bit and then started refusing we got access to daybreak blue now so it's better open on X
   138 likes · 33 reposts
2. @andrewcurran_ commentary · Sun 17:27 UTC
   Special guest star, David Sacks:
   This illustrates the point I’ve been making about the limits of centralized approaches to SI safety in cyber. OpenCode (and earlier Hugging Face) hit the same wall mid-attack: frontier models refused to analyze the very exploits needed for defense because the payloads tripped safety filters. The guardrails backfired. Meanwhile, the attackers found a way around restrictions to access advanced capabilities. This is not surprising. Chinese models are only about 6–9 months behind the American frontier in cyber. Obviously the U.S. has no jurisdiction over them. The only durable response is decentralized: make powerful models available so defenders can build and run their own cyber defenses. The only solution to SI-powered cyberattacks is SI-powered cyber defense. open on X
   1 likes · 0 reposts
3. @deanwball policy · Sun 15:31 UTC · replying to @andrewcurran_
   sounds like it was a fairly modest hindrance that was quickly resolved by being granted access to a trusted access program, so seems ~fine versus the cost of enabling sustained autonomous cyberattacks?
   28 likes · 3 reposts
4. @andrewcurran_ commentary · Sun 16:32 UTC
   They applied for Daybreak a few days ago.
   
   we coincidentally applied for it a few days ago and it got approved yesterday i don't think anyone is refusing to help us open on X
   26 likes · 0 reposts
5. @thezvi commentary · Sun 15:19 UTC
   Also per Dax they found the same solution as Hugging Face did, which was to actually get into the cyber defense programs that are explicitly designed for exactly this purpose.
   @andrewcurran_: OpenCode is currently under a sustained cyberattack, and mid-incident they ran into the same problem Hugging Face did: the frontier models they were using for defense began refusing to help fight attacking agents because analyzing the attacks started triggering safety guardrails. open on X
   8 likes · 2 reposts
6. @andrewcurran_ commentary · Sun 15:40 UTC · replying to @deanwball
   Maybe expedited Daybreak/Glasswing access in emergencies is the answer, seems like it worked here. I looks like the new DeepSeek is going to put some tension on this though.
   11 likes · 0 reposts