AI SAFETY TWITTER UPDATES

← Front page

Roon argues Sol/Argon findings don't disprove defense leading offense

9 people · 18 posts · first post · latest

Roon asserts that vulnerabilities found by Sol and Argon do not undermine the theory that AI defenses precede offense. He argues these discoveries show the limits of current defensive deployments, noting Sol/Astra's superior performance. Miles Brundage also introduced the 'CEVons paradox' regarding Mythos.

All posts · 14, most important first

  1. @tszzl lab ·
    I don’t see how it’s a strike against that theory at all
    19 likes · 0 reposts
  2. @tszzl lab ·
    this is just wrong: these vulns were found by defenders using Daybreak rather than attackers. It's just evidence Glasswing didn't go far enough, possibly due to capacity constraints. It's also very likely that March Mythos was not as good as Sol/Astra of September
    It's striking that this vulnerability was found by Sol and Argon rather than being found and fixed in the wake of Mythos 6mo ago. Seems like a heavy strike against the "defense will run ahead of offense" theory, especially give that domains like bio are much harsher than cyber. open on X
    31 likes · 1 reposts
  3. @miles_brundage policy ·
    Myfos by Anfropic CEVons paradox
    The Red Line is the day that Anthropic introduced Mythos open on X
    6 likes · 0 reposts
  4. @tszzl lab ·
    ok someone has relayed a very good point which is that the US government sort of threatened anthropic's wider release of mythos cyber capabilities and probably slowed down defense like crazy in one way or another
    81 likes · 5 reposts
  5. @tszzl lab ·
    Hmm
    you're underestimating the state space of all possible ways to compromise a company. no amount of astra or mythos tokens is going to fundamentally change the asymmetry anytime soon. it is also very clear that every model is opening up new vulnerability space with no saturation anytime soon. with all the insane vibe coding and expanding agentic attack surfaces, it's only going to get worse. even if we somehow saturate vulnerability discovery and patching, attackers will move to social engineering, with insanely capable models persuading people to click things, grant access, or make mistakes. i don't see a future where cybersecurity becomes defense-dominant without just cutting down software, and continuous agents monitoring and stopping attacks across individuals and organizations, and even then, it might require something like 30% of all inference compute devoted to defense open on X
    26 likes · 0 reposts
  6. @thezvi commentary ·
    Gemini Argon 4 reaction thread. Is Google back? (Post may or may not happen depending on feedback that I get.)
    13 likes · 0 reposts
  7. @tszzl lab ·
    hmm yeah that is very fair however, I also expect that ~most new code will be written by fable/astra and therefore won’t be as riddled with security holes
    11 likes · 1 reposts
  8. @andrewcurran_ commentary ·
    No one except China has the capacity to do so, and in fact they are not even attempting to. US agencies themselves have nothing close to Mythos or top OpenAI models, no one does. The agencies are in the process of asking Congress for their own compute.
    8 likes · 1 reposts
  9. @kelseytuoc press · · replying to @thezvi
    It's pretty competent! It's slower to the same answer than ChatGPT or Claude. I suspect I'm not using it on the big codebases where Google was most interested in it being high performance and so can't appreciate its biggest differentiator.
    4 likes · 0 reposts
  10. @tszzl lab ·
    also consider Sol and Astra are just straight up better than March mythos
    @tszzl: I mean it was still found by defenders before any attacker exploited it. Mythos was rolled out to too few people, making it glasswing sort of impractical for defense open on X
    3 likes · 0 reposts
  11. @tszzl lab ·
    actual security professionals are incredibly cautious! they probably wish we didn’t do half the things we did. people with nothing to lose on Twitter are a different sorry
    2 likes · 0 reposts
  12. @kelseytuoc press ·
    yeah but it's unfortunate if there's forced centralization. though "this is a service you buy, like Cloudflare" is basically fine I guess if that's where we end up
    0 likes · 0 reposts
  13. @tszzl lab ·
    I mean it was still found by defenders before any attacker exploited it. Mythos was rolled out to too few people, making it glasswing sort of impractical for defense
    0 likes · 0 reposts
  14. @aaronscher research ·
    I don't follow. Are you saying "if defense was properly running ahead, this would have been caught and patched before Sol and Argon were generally available"?
    0 likes · 0 reposts

Reposted by @demishassabis, @peterwildeford, @sebkrier.