AI Agent & MCP Audit — An AI agent and MCP audit reviews systems where a language model can take actions — calling tools, moving funds, editing data — and establishes what an attacker achieves by controlling any text the model reads: prompt injection, tool poisoning, RAG poisoning, privilege escalation and unbounded autonomous action.
An agent with tools is a remote code execution surface with a natural-language parser in front of it. Anything the model reads — a webpage, a document, a transaction memo, a GitHub issue, an MCP tool description — is untrusted input that can redirect its behaviour. The security question is not whether the model can be tricked; it is what the model is permitted to do once it has been.
We audit the blast radius first: every tool the agent can call, every credential it holds, every irreversible action it can take, and every boundary that is enforced by a prompt rather than by code. Then we attack it — direct and indirect injection, tool description poisoning, retrieved-context poisoning, confused-deputy chains across MCP servers — and report what survives.
We fix a commit hash, agree the in-scope contracts and read your architecture docs, then build a threat model: who the actors are, what the trust boundaries are, and which invariants must never break. Nothing is reviewed against assumptions we have not written down.
Line-by-line review by at least two auditors working independently, focused on authorisation, accounting, upgrade paths, external integrations and the gap between what the code does and what the documentation claims it does. Most critical findings come from this phase, not from tooling.
Static analysers appropriate to the language, plus property-based fuzzing and invariant testing to push the system into states no unit test covers. Tooling is used to widen coverage, never to replace the manual pass.
Candidate findings are proven on a forked network with a working proof of concept. We report what an attacker can actually do and what it costs them, not a theoretical severity label.
Every finding gets a severity rating, reproduction steps, the affected code, the impact in concrete terms and a specific remediation. You get a draft for discussion before anything is finalised.
We re-test every remediation against the original proof of concept and check that the fix has not opened a new path. The final report is yours to publish.
| Severity | What it means |
|---|---|
| Critical | Direct loss of funds or permanent freezing of assets, exploitable by any actor. |
| High | Loss of funds or protocol insolvency under realistic conditions, or requiring a privileged actor to misbehave. |
| Medium | Broken protocol behaviour, denial of service, or value leakage that does not directly drain the contract. |
| Low | Edge-case incorrectness with limited impact, or an issue requiring implausible preconditions. |
| Informational | Code quality, gas efficiency, documentation mismatch and defence-in-depth suggestions. |
Single token contract: starts from $999, report in 24–48 hours. dApp, GameFi or RWA project: starts from $2,999. DeFi protocol, L2 / rollup, Bridge, ZK circuit, AI agent / MCP: scoped per project after we have seen the code.
A review of Model Context Protocol servers and the agents that use them: what each server exposes, how it authenticates callers, whether tool descriptions can inject instructions, whether one server can be used to reach another, and what an attacker achieves by controlling any content the model reads.
Not eliminated by prompting — it is an input-trust problem, not a wording problem. It is contained by architecture: least-privilege tools, human gates on irreversible actions, provenance-aware context, output validation and spend limits. We audit whether those controls exist and hold.
Yes, and those get the strictest treatment: signing authority, spend caps, allowlists, slippage bounds and the exact sequence by which a hostile prompt could produce a signed transaction. This is the highest-risk agent category in Web3.
The model boundary. We treat every input the model reads as attacker-controlled and test whether tool descriptions, retrieved documents, memory entries and inter-server calls can redirect behaviour — categories a traditional application pentest does not cover.
Yes: transport and origin validation, session handling, filesystem and network scope, secret storage, and what a malicious client or a malicious server on the same host can reach.
A repository or contract address, a commit hash to freeze the scope, whatever architecture or spec documentation exists, and a point of contact who can answer design questions. If documentation is thin we will write our understanding of the system back to you and ask you to confirm it — that step alone catches design-level bugs.
A single token contract is 24–48 hours. A typical dApp or mid-sized protocol runs one to two weeks. Large DeFi systems, L2s, bridges and ZK circuits are scoped per project after we have seen the code. We will give you a fixed timeline with the quote, not an estimate that moves.
Yes. Fix review is part of the engagement, not an upsell. We re-run the original proof of concept against your patched code and confirm the fix has not introduced a new path.
Send the repository and a commit hash through the contact form, message @bugtester25 on Telegram, or book a 30-minute scoping call. 200+ protocols audited · $4B+ secured · 0 hacks post-audit. Prefer email? info@safeedges.in.