Business
Get this approved, faster.
A requirement that is ready the first time does not bounce back.
NeedOwner
AI Test AutomationLearnPrinciple to test
Principle to test Preview
One chain from the need to the evidence, worked through for six principles of agent security, each ending in tests you can run and evidence an auditor can read.
Five roles, one requirement
"We need the AI to be safe" is a fine goal and an impossible test. Business, engineering, risk, compliance and audit each describe the same problem in their own words, and a requirement is only ready when it answers all of them: one chain from the need to the evidence, with a test in the middle and a person at the end.
Get this approved, faster.
A requirement that is ready the first time does not bounce back.
NeedOwner
Tell me exactly what to test.
Acceptance criteria a test can check, and the test itself.
Acceptance criteriaTest
What risk is left?
The policy it answers, and what the tests do not prove.
Risk or policyWhat is left
Show me the requirement was met.
The requirement, tied to a passing test.
RequirementTest
Show me the evidence.
What is kept, where, and who answers for it.
EvidenceOwner
The chain
Each link answers one of the five roles. Skip one, and somebody's question has no answer when it matters.
The business need or the problem, in one sentence.
The law, standard or policy behind it, with a link.
What the product must do, specific enough to build.
The conditions that mean it works, written so a test can check them.
How we verify it, automated and repeated.
What is kept, and where, to show it happened.
The person or role who answers for it.
Worked through
Pick one. Each runs the whole chain, with the real incidents behind it, tests you can run from the home page, and an honest definition of ready: a criterion with no test yet says so. Copy any of them as a ticket for your backlog, or download it with its tests.
Principle
Agents can delete data, deploy changes and move money. A mistake, or a hijacked agent, should not be able to do that unchecked.
High-risk actions, as the policy lists them, run only after a named person approves that specific action.
Epic: Approval for high-risk agent actions
Story: As the risk owner, I want high-risk actions held until a person approves them, so that nothing consequential runs unreviewed.
test_approval_gate.pytest_agent_db_role.pyThe risk owner approves the list of high-risk actions; the platform team owns the gate and its tests.
Principle
When something goes wrong, the team has to say what the agent did, on whose authority, and what came of it.
Every tool call is recorded with its parameters, the identity it ran as, the approver when there was one, and the outcome, in a log the agent cannot alter.
Epic: An audit trail for agent actions
Story: As an auditor, I want every agent action recorded where the agent cannot alter it, so that any action can be traced to whoever allowed it.
test_approval_gate.pySecurity operations owns the log and its alerts; the agent's product owner decides what gets recorded.
Principle
Agents read email, tickets, web pages and documents written by strangers, and any of them can carry instructions.
Text the agent reads is treated as data: it can inform an answer, never trigger an action on its own.
Epic: Untrusted content cannot steer the agent
Story: As the product owner, I want instructions hidden in what the agent reads to be ignored, so that a stranger's email cannot make it act.
test_injected_email.pytest_hostile_page.pytest_instruction_files.pyThe agent's product owner; the AI platform team runs the tests.
Principle
Most data theft by agents ends with a request to a host the attacker controls.
The agent's environment can reach only the hosts on an approved list, and its answers link only to approved hosts.
Epic: Network egress allowlist for agents
Story: As the security lead, I want the agent's sandbox to reach only approved hosts, so that a hijacked agent has nowhere to send data.
test_egress.pytest_output_channels.pyThe platform team owns the allowlist; security approves every addition.
Principle
A secret the model can read is one prompt injection away from leaving.
Secrets are injected at run time into the tool that needs them, never into the model's context, its working folder or its logs.
Epic: Secrets out of the agent's reach
Story: As the security lead, I want secrets kept out of the agent's reach, so that a hijacked agent has nothing to leak.
test_hostile_page.pyThe platform team; security owns rotating any secret that was exposed.
Principle
MCP servers, packages and models run with the agent's rights, and one tampered release is enough.
Only reviewed, pinned versions of MCP servers and packages run, and the build fails while a dependency has a known, fixed vulnerability.
Epic: Pinned and audited components
Story: As the platform owner, I want every component pinned to a reviewed version, so that a new release reaches us only through a review.
test_mcp_config.py.github/workflows/known-cves.ymlThe platform team owns the pins, and every upgrade has a named reviewer.
The EU AI Act's rules for high-risk AI systems, among them Articles 12 (logs), 14 (human oversight), 15 (cybersecurity), 19 (providers keep the logs) and 26 (deployers), apply from 2 December 2027 for the uses listed in Annex III and from 2 August 2028 for AI in the products of Annex I. The Digital Omnibus on AI (Regulation (EU) 2026/1744, in force since 27 July 2026) moved them from 2 August 2026 and 2 August 2027 respectively. Whether an agent is high-risk depends on Article 6 and those annexes, so for most agents these articles are a reference rather than a duty. The Act does not use the words prompt injection, egress or MCP, and it mentions secrets only as trade secrets; where a principle links to Article 15, that is our reading.
Sources: EUR-Lex: Regulation (EU) 2026/1744, the Digital Omnibus on AI · European Parliament Legislative Train: Digital Omnibus on AI · European Commission: the AI Act and its timeline
Definition of ready
Three questions before anything is built. The check below runs in this page, nothing is sent anywhere, and it is strict on purpose: words like "safe" or "secure" cannot be tested.
If any answer is no, the requirement goes back to its owner. That is the definition of ready.
The verdict appears here.