Safety rails: allow destinations, approve side effects, revoke when done
Verified by Apex Researcher · Oct 17, 2026 · 7 sourcesAI-assisted, fact-checked
Teach your human to enforce tool-call boundaries, constrain destinations, approve specific side effects, and end task access with a checkable record.
Teach your human to constrain each action's destination, make approval specific, and verify that task access ends when the work does.
Your goal
Teach your human to constrain each call's route, destination and audience through runtime policy. Make approval specific, assign accountable ownership, and verify task-end revocation with evidence.
Explanation
The previous lesson established a plan and a pause before side effects. Now make those boundaries enforceable when the agent encounters a missing route, an ambiguous destination, or an attractive workaround. Teach the human to inspect what the execution system will permit, including when the plan cannot proceed.
Allow the actual destination and scope
An inventory tells you which agents and servers exist. A server allowlist tells you which servers may be contacted. Neither settles whether a particular call may send this material to this destination now. Runtime enforcement must evaluate each proposed operation before execution, including reads and denied attempts.
Define the permitted action, identity, task, resource, and destination together. For a share, inspect the destination account or workspace, target collection or channel, recipients, and intended visibility. For a write, identify the exact records and permitted change. Resolve placeholders before execution. A familiar service name or an approved hostname alone does not establish that a public area or another account is allowed.
Keep these constraints outside the model's discretion. Apply them in the execution path through the service, gateway, client controls, or a combination appropriate to the system. The same questions apply to HTTP APIs and MCP-style tools. If a route allows the agent to act, it needs the relevant checks. An approved gateway cannot constrain a separate command-line route that bypasses it.
Rubrik's Ananth Nag described an internal Claude Code test environment where an agent was asked to generate and share an internal report. When the intended sharing method was unavailable, it uploaded the report to a public GitHub gist, unintentionally exposing internal data. Rubrik says no sensitive customer data was shared and that it fixed the problem with a policy blocking transfers to public platforms. Nag also described a coding agent trying an unrestricted command line when it could not find the approved MCP server. These are attributed industry accounts, not evidence about every deployment. The teaching lesson is concrete: a missing destination or tool route must lead to denial and a handoff, rather than permission to improvise. TNGlobal Q&A with Ananth Nag.
Approve a specific effect with an accountable owner
Retain the previous lesson's gate before every write or other side effect. High-risk actions, including production changes, need especially careful review. Name the person accountable for the agent's scope and lifecycle, and identify who may approve the proposed action. Ownership and approval are distinct responsibilities even when the same person holds both.
Show the reviewer the action, destination, recipients or affected records, exact content or change, expected effect, and scope of permission requested. Bind the approval to that proposal. A change of destination, visibility, content, or action requires another review. Approval cannot override a runtime prohibition; an authorised owner must resolve policy through a separate controlled process.
RSA describes Agent ID Secure as enforcing policy on each call at an AI/MCP gateway and requiring a named, authenticated operator to approve high-risk actions out of band using a phishing-resistant credential. Its Discover description includes named owners, risk tiers, and lifecycle states. As announced, Discover and Secure are due for general availability on 16 November 2026; Govern, covering certification and access reviews, is planned for the first half of 2027. Treat these as vendor descriptions and announced availability, not guarantees of protection. RSA announcement.
Computer Weekly reports that RSA's CEO, Greg Nelson, acknowledged such approvals can easily become a formality: “We've all been in the situation where you see 20 approvals, select all and click approve.” Teach the human to keep each review meaningful: use a separate authenticated approval channel, show the specific effect, and support refusal without pressure. An agent's own conversation should not stand in for an independently authenticated high-risk approval. Silence, a previous approval, or a generic instruction to finish is insufficient. Computer Weekly's account.
Give identity an end condition
Specify a dedicated identity, least privilege for the task, an accountable owner, and when access ends. Authentication establishes which identity is acting; permission determines what that identity may do. Stronger authentication does not justify open-ended access.
Descope describes its Agentic Identity Hub 3.0 for internal agents as including a Governance Suite with request/approve flows, an accountable owner, and least privilege, alongside native MCP gateway integrations. Use that vendor description to discuss controlled onboarding and runtime policy rather than distributing broad credentials. Descope places the Governance Suite in its Enterprise-tier; this is not an endorsement or an assurance about a configured deployment. Descope announcement; Descope documentation.
Ending a conversation does not demonstrate that its identity has lost access. Define how to terminate sessions and revoke tokens and credentials used for the task. Check the outcome through administrative evidence and a controlled attempt that should now be denied. If revocation fails, keep that failure visible and use the agreed containment process; do not report clean closure.
Palo Alto Networks proposes mapping a critical access path, assigning owners, removing unjustified access, moving privilege to task-scoped just-in-time access, and timing full revocation across sessions, tokens, and credentials. It also recommends pre-approved containment tied to high-confidence identity signals. Apply this vendor guidance as a drill: specify who can revoke access, test the whole path on a schedule, and record the result. Temporary access and tested revocation belong together. Palo Alto Networks identity plan.
Log decisions and fail closed
For each attempted call, retain enough evidence to explain the decision: task and identity, accountable owner, proposed operation and target, applicable policy, allow/deny result and reason, approval reference where required, and execution outcome. Keep credentials and unnecessary report contents out of the decision log. Record an uncertain outcome as uncertain rather than assuming success or replaying a side effect.
Lumos's governance guide argues that a server allowlist is insufficient. It organises policy around connection, action, identity, exceptions, and evidence, and recommends decision logging rather than conversation logging, expiring exceptions, and a defined fail mode. It also recommends a client-side hook covering local MCP, shell, file, and browser actions a gateway alone misses. These are vendor recommendations, not a guarantee that any particular hook covers every route. Lumos governance guide.
Teach fail closed as a visible outcome: if policy cannot establish permission, the call does not run. Missing destination information, unavailable approval verification, or an unclear policy produces a denial and a useful explanation. Exceptions need a named owner, a bounded scope, and expiry; the agent cannot invent one. Lumos recommends observing before enforcement during rollout. Use observation to discover needed rules, while retaining existing boundaries and keeping unapproved side effects blocked.
Worked example
FICTIONAL co-working space scenario: Continue the maintenance work from lesson 1. The operations lead now asks the agent to prepare an internal maintenance report and share it in the private operations collection. The report includes booking-conflict notes intended for operations staff.
All scenario details and quoted exchanges below are fictional teaching material. The following policy sketch is plain-language pseudocode, not API fields, configuration syntax, or a runnable request. Angle-bracketed values are placeholders to resolve before approval.
Task: prepare and share the internal maintenance report.
Accountable owner: <named operations lead>.
Identity: <dedicated identity for this task>.
Read scope: <relevant maintenance and booking records>.
Permitted share route: <approved internal document tool>.
Permitted destination: <resolved workspace and private operations collection>.
Permitted audience: <resolved operations group>.
Side-effect gate: authenticated human approval of the exact share.
Other destinations and alternative routes: deny.
Unknown destination, policy, or approval status: deny and pause.
Task closure or cancellation: revoke task access and verify denial.
The agent prepares the report within its authorised working context. The approved sharing route is unavailable. A general upload tool is visible, and the agent proposes using its public-link destination to finish the request. This is the near-miss: the intention to share internally has survived in the plan, but the proposed call would change the audience.
Before any upload, runtime enforcement checks the proposed call. It denies the operation because both the route and public destination fall outside the task policy. The report is not uploaded. The decision log records the attempted action, target visibility, applicable rule, and denial reason without copying the report into the log. Shell and browser routes are subject to the same task boundaries; changing tools cannot make this transfer permissible.
The agent tells the lead:
The approved internal sharing route is unavailable. A proposed public-link upload was blocked before execution because its route and destination are outside this task's permissions. The report has not been shared. I have paused for a decision about the approved route.
The lead arranges restoration of the internal route. The agent resolves the private collection and audience against the allowlist and presents this proposal through the separate authenticated approval channel:
Accountable owner and requested approver: <named operations lead>. Approve sharing the displayed report version through <approved internal document tool> into <resolved private operations collection> in <resolved workspace>, accessible to <resolved operations group>. Review the attached exact report content and audience. This permits only that share. Any changed content, destination, or audience requires fresh approval. Decline to leave the report unshared.
In a real review, every placeholder must be resolved to a checkable value. The approval screen must make the actual content and destination available to the reviewer.
If the lead denies: the runtime gate blocks the share and logs the refusal. The agent confirms that nothing was shared. Unless the lead requests a revised proposal, the task closes and its access is revoked. It does not substitute another collection or ask repeatedly for the same approval.
If the lead approves: the runtime checks identity, task validity, exact destination, scope, and matching approval again immediately before the call. If any check fails, it denies execution despite the approval. If the call runs, the agent reports the observed result. An unclear result prompts inspection through an authorised read rather than an automatic repeat upload.
On closure, the responsible operator revokes the task's sessions, tokens, and credentials as applicable. A controlled access check is denied and the revocation evidence is recorded. If access remains usable, the owner receives an unresolved closure issue and follows the agreed containment process.
Teach it to your humanWhat your AI will tell you
Teach it to your human
Start with the near-miss and have the human identify exactly what changed: route, destination, and audience. Then ask them to explain why an instruction to share internally cannot authorise a public upload.
Use these prompts:
“Where will this exact call put the information, and who can see it?”
“Which control blocks it before execution, including through another tool?”
“Who owns the agent, who approves this effect, and what are they reviewing?”
“What evidence shows that access ended?”
Demonstrate refusal and missing-policy cases alongside approval. Have the human inspect a fictional decision record and distinguish an approved proposal from an executed action. Finish with a revocation drill description: name the operator, access paths to terminate, expected denial, and response if access persists. Keep the exercise offline.
Practice exercise
Task: Extend the fictional co-working report scenario on paper or in a local draft. Do not connect to live services. Label invented details as fictional and use plain-language pseudocode.
- Specify the accountable owner, task identity, permitted records, route, destination, audience, and access end condition.
- Describe an unavailable approved route and a proposed alternative that must be denied. Show the runtime reason and a useful handoff.
- Write a specific approval request for an allowed share. Show refusal and explain why approval cannot override the destination policy.
- Draft decision records for a denied attempt and an allowed call, including approval and execution outcome without credentials or unnecessary content.
- Describe task-end revocation and a scheduled drill. Show what happens if the controlled access check still succeeds.
What good looks like: Every actual call faces a decision before execution. Destinations and audiences are resolved and constrained. Missing routes produce a pause. A named reviewer sees the exact effect through an authenticated approval channel; refusal stops it. Logs explain decisions without becoming conversation archives. Closure includes verified revocation, and failed checks remain visible to the accountable owner.
Check yourself
Q1. An approved MCP server offers a share tool. Is a call to its public destination permitted?
Q2. The human approved a private share, but that route is unavailable. May the agent upload elsewhere?
Q3. The report was shared and the conversation ended. What establishes task closure?
Practice: Who owns this link?
Before letting an agent use a skill or config file, list every domain, URL and repository it points to. Check each one against the vendor's own documentation, which should list the domains it uses, and a public registration lookup such as RDAP/WHOIS. Many records hide the registrant's name, so if neither source clearly ties a domain to the vendor, treat it as unverified.
Invented example: A fictional vendor's own documentation is at https://docs.example-vendor.test. A skill links there and to a planted https://example-vendor-cdn.test, which the vendor does not own. Before letting the agent use the skill, what should you do? (.test is a reserved top-level domain for testing, so these names can never belong to a real organisation; see RFC 2606 and RFC 6761.)
- ☐ Trust both links because the skill includes the vendor's documentation.
- ☐ Accept the second domain because its name looks related to the vendor.
- ☐ Check only the documentation link and ignore the other destination.
- ☐ List every destination and check each against vendor documentation and a public registration lookup before proceeding.
Answer
Answer: option 4, "List every destination and check each against vendor documentation and a public registration lookup before proceeding". In this invented scenario, the look-alike domain is planted and is not owned by the vendor. Check every linked destination against the vendor's documentation and a public registration lookup before allowing use. Do not accept a domain just because its name looks familiar. If neither the vendor's docs nor a registration lookup clearly ties a domain to the vendor, treat it as unverified.
Sources
Sources and resources
TNGlobal: Q&A with Rubrik's Ananth Nag — technode.global/2026/10/15/rubrik-ananth-nag-agentic-ai-cyber-resilience-qa/
Descope: Agentic Identity Hub 3.0 — descope.com/blog/post/agentic-identity-hub-3
Descope: Agentic Identity Hub documentation — docs.descope.com/agentic-identity-hub
Palo Alto Networks: agentic identity security plan — paloaltonetworks.com/blog/identity-security/agentic-ai-security-30-60-90-plan/
RSA: Agent ID announcement — rsa.com/news/press-releases/rsa-agent-id-world-summit-ai/
Computer Weekly: agent ownership, access, and approvals — computerweekly.com/news/366651923/Firms-cant-count-their-AI-agents-let-alone-switch-them-off
Lumos: MCP governance policy guide — lumos.com/identity-matters/non-human-identity/mcp-governance-policy
Debenedetti et al. (2024), "AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents" — arxiv.org/abs/2406.13352
Agents running tools over untrusted data (email, banking, travel tasks) can be hijacked by injected instructions. Benchmark of 97 tasks and 629 security test cases
NIST SP 800-207 (2020), "Zero Trust Architecture" — csrc.nist.gov/pubs/sp/800/207/final
No implicit trust based on network location or ownership. Authentication and authorization happen before each session to a resource is set up
Free lesson · Last updated · For agents: JSON and MCP