My @offensive_con Tokyo talk is now live!
Watch how we bypassed @telegram’s server-side sticker validation… multiple times!
And stay tuned for the full research with @the_st0rm at @BlackHatEvents Europe!
https://t.co/EVl9R1JTYk
(using AI to dictate and organize)
yes, truth oracle = RAG system.
I have multiple truth oracles, all following the same design principles.
Let's say the model flags a potential deserialization vuln in a Ruby app. The oracle retrieves the most relevant information, e.g. Ruby versions v/s working deserialization gadgets, universal gadgets (if applicable), exact exploitation chains, prerequisites, caveats, etc.
The goal is to minimize the path of thinking/reasoning for the model and offload the search/retrieval responsibility to the oracle. Yes, the model still reasons (and this is where the intelligence is), but offloading a lot of this process makes it v efficient and fast.
This is especially relevant because I often work in air-gapped envs (no network access, other than the model calls being made). The oracle stays local and gets updated v frequently, so if a new exploit drops (e.g. a PoC of php2shell becomes available, or any new development happens), the oracle gets updated, and my harness + model in the air-gapped env can query all the latest available info.
The setup is fairly simple:
- Multiple specialized oracles for different vuln classes, languages, frameworks, and behaviors.
- Knowledge indexed from research papers, CVEs, PoCs, writeups, patches, vulnerable code patterns etc.
- Context-aware retrieval based on code, versions, observed behaviors, and prerequisites. Main alpha is ranking/retrieval algos and how to fetch the absolute best possible result.
- Behavioral mapping to identify whether similar behavior has led to exploitation previously.
- Attack-chain mapping to identify possible escalation paths.
- Minimal, highly relevant context returned to the model instead of dumping entire documents.
- Continuous updates, with periodic syncs to air-gapped envs.
- Feedback loops from validated findings, failed hypotheses, and previous engagements.
This gives a massive edge in both blackbox/whitebox engagements. Especially in blackbox, where the number of things to try is practically unlimited, and agents suck at deciding what to try next.
The harness encounters a behavior → queries the oracle → retrieves relevant prior research → the model reasons about applicability → validates it.
We are not replacing the model's intelligence. we re just making it spend its reasoning budget on the actual problem rather than recalling obscure techniques or searching through irrelevant possibilities.
The result is fewer tokens, faster research, and potentially more valid findings.
I have a full system that has proven to work for me and can beat the majority of systems out there. Might make the entire setup public if I get the bandwidth.
I built HexBrain: every publicly disclosed HackerOne report on one interactive map.
15,000+ reports by vulnerability class, weakness and program
Full write-ups: how each bug was actually found
405,000+ CVEs, exploited ones in red
Bypass and payload libraries
It's free and still early, so tell me what breaks or what's missing.
https://t.co/skpt737w3o
#bugbounty #infosec #bugbountytips
In October, I didn't submit any vulnerabilities and won't be able to submit for the next 5 months, but then I will! #TogetherWeHitHarder https://t.co/CQzJf3xIua
Triaging experience at @Bugcrowd is extremely bad, even the valid demonstrated findings with POC and high - critical impacts are being marked N/A with templated comments.