Kyvvu

OWASP agentic · ASI08

ASI08: Cascading Failures

A single fault, such as a hallucination, injected input, faulty tool or poisoned memory, spreads across connected agents and workflows and grows into system-wide harm. Because agents plan, delegate and persist state without step-by-step human checks, errors can propagate faster than people can respond. This entry is about propagation and amplification; the original flaw is classified under its own entry.

What it looks like

  • A compromised planning agent emits unsafe steps that executor agents carry out automatically.
  • One poisoned update or false alert makes peer agents take disruptive action across regions.
  • Two agents rely on each other's outputs and reinforce an initial error in a feedback loop.

What OWASP recommends

  • Design for failure: isolate agents with trust boundaries, least privilege and segmentation to contain spread.
  • Separate planning from execution with an independent policy engine, and validate outputs or add human gates before they propagate.
  • Rate-limit and add circuit breakers and blast-radius caps that pause on fast-spreading anomalies.
  • Keep tamper-evident logs of inter-agent messages and decisions with lineage, so cascades can be traced and rolled back.

How Kyvvu's policies address it

These policies from the open library are written for this threat. Runtime policies judge each step against everything the agent did earlier in the task; registration policies check what an agent declares before it runs.

Based on the OWASP Top 10 for Agentic Applications for 2026 by the OWASP GenAI Security Project, licensed CC BY-SA 4.0. The names are OWASP's; the summaries are ours.

An example policy

Bound resource calls per task to prevent runaway loops

From the OWASP agentic default pack. Severity high: the step runs and an incident is opened.

- name: "Bound resource calls per task to prevent runaway loops"  description: |    ASI08 (Cascading Failures), ASI02 (Tool Misuse - looping pattern):    Blocks tasks that issue more than 50 step.resource calls.    Catches the most common dev-time embarrassment: an agent stuck in    a tool-call loop, which under OWASP's framing is both a DoS/cost    issue and a precursor to cascading failures across connected agents.    50 is generous enough that legitimate multi-step workflows do not    hit it; production deployments typically tighten per-agent based    on observed normal behaviour.    Applies to: ALL agents.  rule_type: execution_max_steps  params:    kind: step.resource    max_steps: 50  severity: high  enforcement_point: step_execution

Your agents

Is ASI08 a risk for yours?

For security teams

Request a demo

We walk through your agents' riskiest paths, and the policies that close them.

For developers

Build the path yourself

Recreate the attack in the playground and see which policies stop it.