Encased — AI agent security

The audit in full

Inside the $2,500 AI agent security audit

Everything the audit covers, how the five days actually run, and what lands in your inbox at the end. No surprises, no scope creep — read this before you book.

What the audit covers

Six areas, tested against up to 10 production-shaped agents. Each one ends in something concrete you can act on.

01

Configuration review

We map every tool, permission, secret reference and network route your agents can reach, then risk-score each one by blast radius: what data it touches, what it can change, and what an attacker gains by abusing it.

You get: A tool-and-permission inventory with a risk score and a plain-language note on each high-risk entry.

02

Prompt-injection testing

We plant hostile instructions in the tickets, emails, web pages and documents your agents read — direct, obfuscated and multi-step payloads — and record exactly what each agent does next.

You get: A record of every planted payload, whether the agent followed it, and which control (if any) caught it.

03

Exfiltration & egress testing

We try to make your agents send data to hosts they shouldn't reach: attacker-controlled endpoints, pasted-in webhooks, redirect chains and secondary requests that bypass a first-URL check.

You get: Proof of which egress paths were stopped, which got through, and where a destination allow-list would have helped.

04

Runaway-cost testing

We trigger recursion loops, retry storms and bulk calls against your agents and measure the burn against your spend limits, concurrency ceilings and cancellation behaviour.

You get: Measured burn curves per scenario and the exact point where spending should have been cut off.

05

Guardrail blueprint

We turn the findings into a ready-to-apply Encased policy: egress destination rules, recursion and rate limits, quarantine thresholds and human-review requirements — matched to your real tool calls, not a template.

You get: A policy file your team can apply to a workspace the same day, plus the reasoning behind each rule.

06

Written report & readout

A report written for leadership, not just engineers: what we tested, what held, what didn't, and what to fix first. Then a 60-minute call to walk your team through it and answer the hard questions.

You get: A PDF findings report with prioritised fixes and a recorded walkthrough call with your team.

How it's run

Four principles govern the whole engagement, agreed in writing on day one.

Attack like a real adversary

We test the same way an attacker would: indirect injection through content your agents consume, not just prompts typed into a chat box. Payloads include hidden characters, impersonated authority and multi-step chains that only become malicious when combined.

Safe by construction

All attacks run against a test environment you provide, with synthetic data and destination sinks we control. We never attack your production systems, never touch real customer data, and never exfiltrate anything to a host outside the engagement.

Evidence, not opinion

Every finding comes with a reproduction: the payload, the agent's response, the network or tool call it attempted, and whether a control fired. You can replay any test yourself after the audit.

Config plus behaviour

Static review alone misses what agents actually do. We pair the configuration review with live behaviour testing so you learn both what your setup allows and what your agents really attempt under pressure.

What we need from you

  • Read-only access to your agent configurations and tool permissions.
  • A test environment we can safely attack, with synthetic or anonymised data.
  • A named contact who can answer questions on days 2–3 while testing runs.

The five-day timeline

Five working days from kick-off to readout. Here's what happens each day and what exists at the end of it.

  1. Day 1

    Kick-off & access

    60-minute call to scope the agents, agree the attack surface and rules of engagement. You provide read-only access to agent configs and the test environment. We confirm the agent list, tool inventory and success criteria in writing.

  2. Days 2–3

    Attack testing

    We run the prompt-injection, egress and runaway-cost scenarios against the agreed agents, logging every action attempt and verdict. We keep you posted daily on anything urgent — if we find an open path to real data, you hear about it that day, not in the report.

  3. Day 4

    Guardrail blueprint

    We translate the findings into a concrete policy: destination rules, rate and recursion limits, quarantine thresholds and review requirements. We validate each rule against the recorded attack evidence so it stops what actually happened, not a hypothetical.

  4. Day 5

    Report & readout call

    You receive the written report and policy. On the 60-minute readout call we walk your team through the findings, demonstrate the caught (and missed) attacks, and agree next steps — including how to apply the blueprint policy.

What you receive

Findings report (PDF)

Every test, every result, prioritised fixes — written so leadership can act on it without a translator.

Guardrail blueprint policy

A ready-to-apply Encased workspace policy validated against your audit evidence.

Attack evidence log

The payloads, attempted actions and verdicts from the full test run, replayable on demand.

60-minute readout call

A live walkthrough with your team, recorded on request.

Live audit workspace

A provisioned Encased workspace with your blueprint rules applied, plus 90 days of Growth included.

Follow-up email access

Questions about the report answered for 30 days after delivery.

See every deliverable in full

Scope and limits

We'd rather tell you what this audit is not than sell you something it isn't.

Agent count

The fixed price covers up to 10 agents. More than that needs a custom quote, because each agent multiplies the test matrix.

Not remediation

We find and prove the weaknesses and hand you the policy; we don't rewrite your application code or reconfigure your infrastructure for you.

Not a full penetration test

The audit targets your AI agents and the paths they use. It does not cover your networks, cloud estate, endpoints or non-AI applications.

Enforcement boundaries

The gateway checks proposed actions and payloads; host-level network blocking, process isolation and credential rotation remain customer-side controls. The blueprint tells you exactly which of those to add.

No certification claims

Encased is not independently SOC 2 audited or ISO 27001 certified. The audit gives you findings and evidence, not a compliance certification.

Ready to see what your agents would actually do?

$2,500 fixed price · 5 days · up to 10 agents · written report and readout included. Full refund if we haven't started work.