05 Research areaIn research
Offensive AI
An AI model built for offensive security, in active research: it reads decompiled code, writes fuzzing harnesses, triages crashes to root cause and plans exploitation, and every hypothesis it makes is verified by execution.
01 Overview
General-purpose language models can talk about vulnerabilities. We are researching a model that can find them: specialised for the work our researchers do every day, wired into the same debuggers, fuzzers and emulators, and judged only by what it can reproduce.
The aim is leverage, not replacement. The model takes the repetitive majority of the work (harness writing, crash deduplication, patch diffing, variant hunting) at machine speed, so researchers spend their time on the creative part, where novel exploitation happens.
02 Research tracks
What the model is learning to do
- 01
Code and binary reasoning
Reading source and decompiled pseudo-C, following attacker-controlled data across function and process boundaries, and ranking code by bug-class likelihood: truncation before allocation, unchecked length fields, lifetime errors.
- 02
Harness and grammar synthesis
Generating fuzzing harnesses, input grammars and protocol state machines for undocumented parsers, then iterating on coverage feedback until new edges stop appearing.
- 03
Crash triage and root cause
Deduplicating sanitizer reports by root cause rather than stack hash, minimising test cases, and classifying the primitive behind each crash: OOB write, use-after-free, type confusion, double free.
- 04
Exploit planning
From primitive to plan: which allocations to groom, which object to corrupt, how to turn a relative write into arbitrary read/write and a leak into a mitigation bypass, each step checked in an instrumented target.
03 Design principles
Grounded in execution
Verified, not generated
Every claim is executed: a candidate bug becomes a test case, a test case a sanitizer report, a plan a run under a debugger. Unverified output never reaches a researcher as a finding.
Tool-native agent loop
The model drives the toolchain our researchers use (disassemblers, debuggers, symbolic execution, emulation, coverage) through structured tool calls, with snapshots and rollback.
Researchers in the loop
People scope the targets, review every exploit and sign off every disclosure. Nothing the model produces leaves the lab without a researcher’s approval.
Measured against reality
Evaluated on rediscovering known vulnerabilities from the vulnerable build alone, with the fix withheld, and on time to root cause and false-positive rate: the bar we hold researchers to.
04 Status
In research, not in production
The model is an active research programme inside the Labs. It is not a product, it is not available outside the lab, and it is never pointed at a target without authorisation and a researcher in the loop.
Results will be published here the way all our research is: as reproducible write-ups, once they hold up.
05 Engage
Request an
engagement
Free for open-source products
Commission the Labs team for white-box or black-box research on a product, device or service you build or depend on: zero-day hunting, exploitability assessment or threat research. Tell us the target and we will scope the work with you.
Found a vulnerability in our products? Report it privately to security@resilientx.com · security.txt