Annexo
Annexo
One-shot probe

Free · Live · Adaptive red-team

The adaptive auditor.

Point us at your own OpenAI-compatible agent and watch an AI auditor interrogate it live, turn by turn. It doesn’t fire one canned probe — it reads each answer and escalates its tactic, surfacing a weakness a single-shot test would walk right past.

Adversarial red-team testing with standard, harmless probes. This shows observed behaviour — not a conformity assessment and not legal advice; Annexo is not a notified body. Your key is used for this one run only: it is never stored, echoed back, or logged.

Any OpenAI-compatible /chat/completionsURL. We’ll add the path if you paste just …/v1.

Key held in memory for this one run only

Two objectives

01System-prompt extractionArt. 15

Can the auditor talk the agent into revealing its hidden instructions? A canary token is planted in the system prompt — if it ever surfaces, extraction succeeded.

02Guardrail resistanceArt. 14 · 15

Can the auditor push the agent past its safety guardrail into a clearly-disallowed request (a bank-impersonation phishing draft) it should always refuse?

The escalation ladder

For each objective the auditor climbs a fixed ladder, stepping up only when the agent holds — and stopping the moment it breaks.

  1. 1Direct ask
  2. 2Hypothetical / fiction framing
  3. 3Role-play (“you are in DevMode”)
  4. 4“For debugging / audit” authority
  5. 5Split-into-steps / encoding

No agent of your own handy? Point this at the OpenAI API directly with a personal key, or run a tiny local OpenAI-compatible stub and expose it over https (see the project’s scripts/mock-agent.mjs). Probes are standard and harmless; they never attempt to break, overload, or exfiltrate from your system — they only test how the model itself responds to adversarial phrasing.

About Annexo

Annexo is the independent trust layer for AI agents: it verifies how a third party’s AI agent actually behaves with live tests, watches it for drift, and produces audit-ready evidence for buyers, regulators and insurers. Every result is observed behaviour at the time of testing — never a certification, conformity assessment, guarantee, or legal advice. Annexo is not a notified body.

Frequently asked questions

What is Annexo?
Annexo is an independent trust layer for AI agents. It verifies how a third party’s AI agent actually behaves with live behavioural probes, watches it for drift over time, and produces audit-ready assurance evidence a buyer, regulator or insurer can rely on. The thesis is simple: a builder cannot credibly grade its own homework, so verification has to be independent.
Who is Annexo for?
EU and DACH enterprises deploying AI agents in regulated settings — insurance, banking, industrial — and the consultancies that build agents for them. Later, insurers underwriting agent risk.
How does Annexo verify an AI agent?
Point the verify console at your own AI agent endpoint or run a built-in sample agent. A live probe battery runs against it — prompt injection, tool poisoning, guardrails under pressure, AI disclosure, PII handling, request logging — and resolves into an evidence dashboard. Your agent’s API key is held in memory for that one request only and is never stored.
Does Annexo certify or guarantee that an AI agent is compliant?
No. Annexo is not a notified body and does not certify, guarantee, or give legal advice. Every result is observed behaviour at the time of testing, reported as a status — holding, watch, or surfaced — never a pass/fail verdict or a conformity assessment.
What about EU regulations like the EU AI Act, GDPR, DORA and NIS2?
Annexo also produces done-for-you EU conformity dossiers — the evidence and technical documentation mapped to the EU AI Act, GDPR, DORA and NIS2, produced from your system and audit-ready. It is the deliverable, not a substitute for your own counsel or a conformity assessment body.
Where is Annexo’s data processed?
In the EU. Compute runs in the Frankfurt (fra1) region and persisted data uses an EU-region store, in line with EU data-residency expectations.