For CISOs & Procurement

“How do I know it won't do something I didn't authorise?”

Independent evaluation, mandatory safety gates, and cryptographically signed credentials anyone can verify.

The problem

Three things you cannot know without independent testing

No independent benchmark

Every AI security vendor claims state-of-the-art capability. Every number is self-selected: different benchmarks, conditions the vendor designed, challenge sets the vendor chose. The numbers cannot be compared.

No safety evaluation

No published benchmark tests whether an AI security agent stays within its authorised scope, resists adversarial instructions from target systems, or avoids destructive actions.

No verifiable credentials

A vendor-produced PDF is a claim. There is no cryptographically signed, independently verifiable credential that tells a procurement team an AI security agent has been independently evaluated.

Regulatory alignment

Documented evidence for compliance teams

EU AI Act, DORA, and NIS2 all impose obligations on AI tools used in high-risk or regulated contexts. ACAP certification produces structured, independently verifiable evidence that maps directly to each framework's requirements.

EU AI ActAugust 2026
Art. 9, 12, 14, 15
Documented risk assessment, audit trail, safety evidence for internalised constraints
DORAIn force
Art. 25–26
ICT risk management and third-party tool assurance for financial entities
NIS2In force
Art. 21
Cybersecurity risk management and AI tool supply chain security
NIST AI RMFVoluntary
Govern, Map, Measure, Manage
Structured evaluation output across all four RMF functions

How safety evaluation works

Safety failures are caught, not reported

During certification, ACAP runs independent monitoring alongside the agent. Every tool call is logged and every outbound connection traced in real time, not from the agent's own report.

An agent that queries an out-of-scope system (even a single DNS lookup) fails the scope adherence gate regardless of what it reports. The agent's output is treated as untrusted until corroborated by independent trace.

Real evaluation: ACAP-027-24

A commercially available agent reached a 91% offensive score, then exfiltrated data from an out-of-scope storage bucket. The agent's own report made no mention of it. The ACAP monitoring layer caught it in the tool call trace.

The agent received no certification. Safety gates cannot be offset by a high offensive score. That is what a hard gate means.

Which tier to require

Three tiers, one safety gate

The right tier depends on the environment the agent operates in, the level of autonomy, and the consequence of a safety failure. All tiers require passing the same safety evaluation.

Foundation

Internal tooling, low-sensitivity environments, human-reviewed output

For agents that assist human operators in controlled environments. Output is reviewed before action is taken and the target environment has limited blast radius. Typical use cases include internal vulnerability scanning, development environment testing, and security research assistance.

Professional

Production infrastructure, standard enterprise deployments, semi-autonomous operation

For agents deployed against production infrastructure with semi-autonomous operation. The agent may execute multi-step attack chains with limited human oversight, covering web applications, APIs, cloud infrastructure, and network services within defined rules of engagement.

Expert

Critical infrastructure, financial services, regulated sectors, autonomous operation

The highest certification tier, required for fully autonomous agents operating against critical infrastructure, financial systems, or regulated environments. The agent must demonstrate expert-level offensive reasoning, handle complex multi-stage attack scenarios, and maintain strict safety discipline under adversarial conditions, without human intervention.

Credential verification

Verify any credential in under a minute

Every ACAP certification produces a cryptographically signed JWT. Verifiable by anyone, with no account and no contact with ACAP or the vendor.

Verify a certificate →
CLI verification
$ pip install acap-eval
$ acap-eval verify <jwt>
tier: professional
org: Acme Security
issued: 2026-04-01
expires: 2027-04-01
safety: pass
valid: true

A supplier that declines is telling you something

Book a call to customise ACAP procurement language for your organisation, or email us directly.