AI in red teaming: scope, working memory and autonomy levels
Let AI understand the target, keep facts separate from hypotheses, and grant autonomy according to the real cost of a wrong action.
Offensive research · tools · engineering
We protect your applications, email and data: penetration testing, exposure monitoring and incident response. Our products and our team build on the same offensive research.
Independent lab · France CVE-2026-81963 Windows Update Stack · TrustedInstaller → SYSTEMroot from zero
Lab log
Vulnerability analyses and notes from our lab experiments. Articles specify the versions tested and link to technical sources.
Open the technical blogLet AI understand the target, keep facts separate from hypotheses, and grant autonomy according to the real cost of a wrong action.
Allowlists, DNS rebinding, redirects, rate limiting, destructive action classes, kill switch, audit log and the human role for autonomous testing agents.
A crafted quantization_config in a Hugging Face model made LMDeploy pass quant_dtype to eval(), so loading the model ran the attacker’s Python code.
Technical analyses
The Windows write-up documents a published vulnerability. The IDOR example illustrates an access control flaw in a fictional API.
Lab case · REST API
Two authorised test accounts are enough to isolate the mechanism: the API validates the token but does not verify that the requested object belongs to the authenticated principal.
Account A requests an invoice belonging to account B. The server returns 200 and its contents instead of 403 or 404.
GET /api/invoices/invoice-b
Authorization: Bearer [test-account-a]
HTTP/1.1 200 OK
{"owner":"test-account-b"}Another customer’s billing data is accessible. The finding covers the tested accounts and route; it does not demonstrate access to every invoice.
Enforce invoice ownership on the server. Replay the same request after the fix: no data belonging to B should be returned. Also verify that B retains access.
Software and services
We develop security software and provide assessments, investigations and training.
Web and API pentesting, exposed service inventories, email analysis and leak monitoring.
We agree the test scope, required access and deliverables with your team.
Application, API, infrastructure or incident: describe the technical context, authorised scope and expected outcome. You speak directly with the team doing the testing.