Security for AI agents and agentic workflows

What we test

• Prompt injection and jailbreaks. Direct and indirect injection through user input, documents, emails, web pages and tool outputs your agent consumes.
• Excessive agency. What your agent is technically allowed to do versus what it should be allowed to do: permissions, scopes and the blast radius of a manipulated agent.
• Data leakage. Whether an attacker can make the agent expose system prompts, training context, other users’ data or internal secrets.
• Tool and integration abuse. The APIs, MCP servers, plugins and function calls your agent can trigger, and whether they can be turned against you.
• Multi-step attack chains. Real attackers combine small weaknesses. We chain findings the way they would, end to end.

How a review runs

1. Step 1Scoping. We map your agents, their tools, data sources and permissions together with your team.
2. Step 2Quote. A detailed quote covering scope, milestones and deliverables.
3. Step 3Testing. Manual adversarial testing by senior researchers. Critical findings are reported as we find them.
4. Step 4Report. Reproduction steps, risk rating and a concrete fix for every finding, written so your developers can act on it.
5. Step 5Unlimited retests. After your fixes are in, we retest until every finding is confirmed closed.

2-4 weeks from kickoff to a full report.