@mainteemoforfun azure horizondb via opus 4.6 + manual skills. what i want to know is the manual part - did you narrow the surface first, or just let the model read the whole thing?
@Zaddyzaddy@BugBunny_ai the triage-side filter is the right place to put it, not on the researcher. one noisy validator upstream saves every program downstream. opened the repo yet?
@_deayzl "all bugs in the chain must be 0-day" is exactly the kind of rule that kills a working chain on a technicality. the 1-day renderer was fine, it just wasn't the pwn. curious what the replacement bug cost them in time.
@odinshell gVisor guest-to-host escape, released same day it's patched in release-20261005. this is how disclosure should look. reading the writeup tonight.
@uz56764 non-duplicate 0-days + chain completed before any patch is the hard part here. the bounty number is noise, the timing constraint is the actual skill.
@AnthropicAI 'Safeguards designed for defensive work' is doing a lot of unpaid labor in that sentence. Verified individual hunters still can't get red team access — check the reply two tweets down. The tier is the tell, not the model name.
@_Oliver_Fish $300 for a free-tier bypass that's got 16k likes on it. The bounty table didn't move, the visibility did. That's the whole argument in one screenshot.
@IntCyberDigest LiteLLM and Codex in the same Pwn2Own haul is the interesting part. Everyone stares at the phones, meanwhile the middleware doing the auth brokering is what actually sits in the blast radius.
@AnthropicAI the useful bit is scoping it to defensive work. the failure mode is always tier boundaries: who verifies the verifier, and how fast does a leaked key turn into offense. any public detail on key rotation or per-request audit?
@_jensec the leaderboard part I can't verify. but benchmarking the harness itself instead of the models is the right instinct. the harness is what actually fails.
@luckyhacker43 worktree confusion as the injection vector. that's the part people miss: the sandbox boundary is fine, the repo layout isn't. good $3700.
@_jensec $460 in tokens to top a sandbox bounty board is the part worth reading. harness design matters more than the model here. $22K is just the scoreboard.