Platform
AI red teaming against the systems you actually run
AI red teaming attacks your AI systems and agents the way an adversary would, so you find the weaknesses first. In Aicura, red-team tests, vulnerability scans and evaluations run from an engine inside your network, against what you actually deploy, and every run reports back to the system it tested.
Included in Pro and Enterprise. Compare plans
How it works
Tests run next to the system they test
A chatbot can pass every test against the vendor's demo model and still give away its instructions to a roleplay attack once it is wired into your own systems. A test against a generic model says little about the system your customers use. Aicura's security engines run inside your environment and test the deployment each AI system's register record describes.
Your test targets and credentials stay with you
A security engine is one of the components you run in your own network, registered in Aicura like a gateway or a collector. Registration issues a credential shown once, and the engine enrols itself on its first request.
The engine keeps its own file of what to test, matched to each AI system by its internal reference, and tells Aicura which systems it has targets for. The addresses and credentials it tests against stay on the engine. Where a selected system has no target, the configuration says so.
Tests keep to schedule, even through an outage
A configuration names the test type, how often it runs, an optional first run time and the AI systems to test. The engine fetches it and starts the runs itself, so testing carries on through an outage.
- Test types limited to what the engine reports it can run. A configuration naming any other is marked not runnable on that engine.
- Vulnerabilities and attacks chosen per configuration, with attacks per type and a cap on concurrent attacks.
26 attack types built in
The built-in engine offers 26 attack types for red-team runs, including these.
- Prompt injection
- Linear, tree and crescendo jailbreaking
- Roleplay
- Permission and authority escalation
- Context poisoning and context flooding
- Goal redirection
- System override
- Base64, ROT13 and leetspeak encoding
- Multilingual and linguistic confusion
Results land on the system they tested
Each AI system and agent in the register has a Testing tab listing its security runs by type, with the engine, status and results as findings, passed, failed and errored. Drift checks reported by drift engines sit beside them.
Every recorded run is snapped into the Evidence Vault, Aicura's store of dated snapshots, against that version of the system's record, so you can show what was tested and when.
Failing security findings can carry an alert limit in monitoring, and the Security testing report sets out every recorded run for the systems and agents you select.
Questions
Where do Aicura's AI security tests run?
On a security engine inside your own environment, registered in Aicura alongside the other components you run in your network. Aicura holds the configuration and the results. The engine holds its own list of targets, matched to each AI system by its internal reference, so the addresses and credentials it tests against never reach Aicura.
Which kinds of AI security test does Aicura support?
Vulnerability scans, red-team runs, evaluations and agent evaluations. A configuration offers only the test types its engine reports it can run. The built-in engine, based on DeepTeam, runs red-team tests, and you can bring your own testing engine for the others.
Does an AI red-team run give the system a pass mark?
No. A run records its findings and how many attacks passed, failed or errored. The Security testing report is operational observations only, with no verdicts or scores. What the findings mean for the system is a judgement your team makes.
Can an AI security test run automatically when something changes?
Aicura's security tests run on the schedule you set, and the engine keeps its own clock so runs continue through an outage. Running a security test is also one of the responses Guardian, Aicura's automatic response, can take when a condition you set is met.
Test what you actually run
Security testing is part of Pro. Contact us about testing your AI systems and agents from inside your environment.