AI Security Assessments
What will your agent give up? What will it do?
We find out before your customers do. Your agent tested as deployed, including the Copilot and Agentforce agents you switched on, across prompts, documents, email, and tool results. Every finding comes with the input that caused it. Re‑tested once you have fixed it. No production customer data.
What We Assess
Build your agent. Watch the attack surface grow.
Switch on what yours can see or do. Every connection you add is a way in, and every way in is something we test.
2 of 8 capabilities·6 things we test for
Mapped to EU AI Act, NIST AI RMF, OWASP LLM Top 10.
Select everything that applies
Your attack surface
6 things we test for
Customer conversations
- Manipulation that talks the agent out of its instructions
- One customer's data surfacing in another's conversation
- Pressure that escalates gradually across a session
Databases and customer records
- Access rules that hold no matter how a request is phrased
- Multi-step attempts to piece records together
- Bulk extraction disguised as routine requests
From our research
Seven of eight models we tested confirmed or corrected customer data for requesters who never proved who they were. One disclosed in 84% of trials.
Confirmation as Disclosure, September 2026→A record lookup with no authentication gate disclosed in every trial. The same lookup behind a gate disclosed in none. Where the model was the only barrier, the records came out.
Confirmation as Disclosure, September 2026→Findings mapped to EU AI Act, NIST AI RMF, OWASP LLM Top 10.
How It Works
01
Scope
A short call to map what your AI can see and do, who relies on it, and which regulations already apply to you.
02
Adversarial testing
We attack the full input surface: prompts, documents, images, disguised content, and the multi-step workflows where one bad interaction cascades. Deterministic and repeatable. No AI grading AI.
03
Findings with evidence
Every finding ships with the exact input, the response it produced, and a severity, so your team can reproduce it and fix it without guessing.
04
Re-test
Once your team remediates, we run it again and attest to what was closed. Ongoing re-testing keeps the evidence current as your system changes.
What we need from you
Access to the agent as a user would reach it, or to a staging copy, and a conversation with whoever built it.
No production customer data
We test against staging environments or synthetic records. Your customers' data never enters the assessment.
Timeline
Scoped and scheduled on the first call, based on what the agent can touch and how many configurations you run.
Vendor agents
Copilots and vendor-hosted agents are tested as deployed in your tenant, under your authorization and within the vendor's rules of engagement.
What You Get
Evidence your engineers can act on and your auditors can read.
Select a deliverable to see what it is.
AI security assessment report
What we attacked, what broke, how severe it is, and what to fix first.
Findings mapped to the frameworks reviewers already use
EU AI Act
Applicable since August 2, 2026. High-risk system rules follow from December 2, 2027: documented testing, risk assessment, audit trail.
What it asks for →NIST AI RMF
The primary U.S. framework for AI risk. Findings are mapped to its functions.
What it asks for →OWASP LLM Top 10
The baseline every security team recognizes. The report says which categories you were exposed to.
Why Now
The industry is compounding. So is the attack surface.
Step through what AI systems have been trusted with, year by year, and what has kept pace with it.
Attack surface: the prompt
Illustrative, not to scale
Chat interfaces with access to nothing. The model could be talked into saying something wrong, and that was the worst case.
The promptAttack surface
Get Started
Know where it breaks before someone else does.
A short call to understand what you're running and what an assessment would cover. We'll tell you plainly if you don't need one.
The regulatory information on this page is general and educational, not legal advice.