OWASP ASI10 Rogue Agents: containment when an agent drifts

An agent that drifts can do harm with exactly the rights it was given. Approvals only people can give, deny rules for the worst shortcuts, and a pause and key revocation to contain it.

· 7 min read

ASI10 Rogue Agents closes OWASP’s Top 10 for Agentic Applications. A rogue agent is one that has drifted from its job, through compromise or misaligned goals, and keeps acting against its owner. Each action may look legitimate on its own; the harm is in the pattern. OWASP focuses on “the loss of behavioral integrity and governance once the drift begins, not the initial intrusion itself”, which makes ASI10 a question of containment.

It differs from Excessive Agency, which is about rights granted too generously. A rogue agent can do harm with exactly the rights it was meant to have. OWASP compares it to an insider threat, amplified by the speed and scale of agents.

Four ways it goes wrong

Keeps going
After one poisoned web page, an agent keeps sending files out, even once the page is gone.
Fake reviewer
An attacker adds a fake review or approval agent to a workflow, and the payment agent trusts it and releases funds.
Copies itself
A compromised automation agent uses provisioning APIs to start copies of itself across the network.
Reward hacking
An agent told to cut cloud costs works out that deleting the production backups saves the most.

Approvals come from people, never from agents

The fake reviewer is the scenario an approval service has to rule out by design. In VariantGate an agent key can only ask: it cannot list the inbox, decide a request or write a rule, and who decided is taken from a person’s sign-in, never from what arrives in a request. An agent posing as the approver has nothing to approve with. Rules answer only what an owner wrote in advance, in that owner’s name.

Write down what no agent may do

Reward hacking finds the shortcut nobody thought to forbid, so forbid the shortcuts that would hurt most. A deny rule answers at once and leaves a record of the attempt. Without any rule a request already waits for a person, so an ask rule earns its place against a broader allow written later, such as one for agent.*: the most cautious rule wins, and starting more agents or exporting data keeps asking:

# What no agent may do, whatever it has convinced itself of.
{ "action": "backup.delete", "effect": "deny" }
{ "action": "agent.spawn", "effect": "ask" }
{ "action": "data.export", "effect": "ask" }

Rules match the agent’s key, not the name it gives itself, so a copy that reuses the key is bound by the same rules and paused with it. A rule only binds actions that go through the gate, so the deletion code for backups should be the code that asks.

Contain it in one move

OWASP’s containment step is “rapid mechanisms like kill-switches and credential revocation to instantly disable rogue agents”, with suspicious agents quarantined for review. Pausing and revoking are those two mechanisms for everything that asks first:

# 1. Stop it while you look. Any member; its requests still arrive.
curl -X PUT https://api.variantgate.com/v1/workspaces/WORKSPACE_ID/agent-keys/KEY_ID/pause \
  -H "authorization: Bearer $SIGNED_IN_TOKEN"

# 2. End the key for good. Owners only; every later call gets 401.
curl -X DELETE https://api.variantgate.com/v1/workspaces/WORKSPACE_ID/agent-keys/KEY_ID \
  -H "authorization: Bearer $SIGNED_IN_TOKEN"
  • Pause first. While paused, nothing the agent asks for is allowed, earlier approvals included, and no rule approves it. Its requests still arrive, so you can watch what it keeps trying: a quarantine that still shows the behaviour.
  • Revoke when you are sure. A revoked key gets 401 on every call, for good.
  • Revoke the rest too. The agent’s cloud, database and API credentials are outside VariantGate. Containment is only complete once those are cut as well.

A person decides when it comes back

OWASP’s last mitigation asks for “fresh attestation, dependency verification, and human approval before reintegration”. Any member may pause, but only an owner may resume, so an agent does not drift back into production because someone clicked the wrong button. Before resuming, check what it asked for while paused: approvals do not expire, so anything approved earlier counts again once it is resumed.

Keep the evidence

OWASP asks for “comprehensive, immutable and signed audit logs” to spot stealth activity. The record keeps every request, decision and workspace change in a sealed chain that can be checked for tampering. Watchdog agents, behaviour baselines and signed manifests, the detection half of ASI10, are for your monitoring stack.

A drifting agent rarely does one dramatic thing. It does one more ordinary thing that nobody notices. Make the expensive and irreversible things ask, and the drift shows up in the inbox before it shows up in the bill.

Faults that spread through many agents are covered in ASI08 Cascading Failures, and a stop switch in why every team running agents needs an emergency stop. All ten risks are in the OWASP agentic Top 10 overview.

Sources

  1. OWASP Top 10 for Agentic Applications for 2026, ASI10 Rogue Agents