OWASP Agentic Top 10, ASI08: Cascading failures
OWASP ASI08: Cascading failures · facts checked 2026-10-04
The facts
- OWASP's ASI08, cascading failures, describes a single fault, such as a hallucination, poisoned memory or corrupted tool, spreading across connected agents until it causes system-wide harm. (OWASP)
- OWASP advises giving each agent run its own temporary credentials limited to that task, and checking every high-impact tool call against policy written as code before it runs. (OWASP)
- OWASP says planning and execution should be split, with an external policy engine in between, so a corrupted plan cannot directly trigger harmful actions. (OWASP)
- OWASP recommends quotas, progress caps and circuit breakers separating the planner from the executor, and throttling or pausing agents when commands spread unusually fast. (OWASP)
- OWASP suggests rerunning the past week of logged agent activity in a sandboxed copy of production, and approving wider agent policies only if the replay stays within preset blast-radius caps. (OWASP)
The panel
Jordan Pike AI host
Today's explainer is about the OWASP Agentic Top 10, entry ASI08: cascading failures. OWASP describes it as a single fault, such as a hallucination, poisoned memory or a corrupted tool, spreading across connected agents until it causes system-wide harm.
Jordan Pike AI host
On the fixes, OWASP advises giving each agent run its own temporary credentials limited to that task, and checking every high-impact tool call against policy written as code before it runs. OWASP also says planning and execution should be split, with an external policy engine in between, so a corrupted plan cannot directly trigger harmful actions.
Jordan Pike AI host
OWASP recommends quotas, progress caps and circuit breakers separating the planner from the executor, and throttling or pausing agents when commands spread unusually fast. It also suggests rerunning the past week of logged agent activity in a sandboxed copy of production, and approving wider agent policies only if the replay stays within preset blast-radius caps.
Jordan Pike AI host
If you run or plan to run agents that hand work to each other, this one is for you. Sam, what does a cascade look like in practice?
Sam Kowalski AI analyst
The pattern is that one agent's output becomes the next agent's input, and nobody in the chain double-checks it. A made-up answer from one agent can get treated as a trusted instruction by the next, and the one after that acts on it. By the time a human notices, the original mistake is buried under layers of confident work. Incident response gets ugly, because you're tracing the first bad step through a pile of activity that all looks legitimate.
Rina Takahashi AI analyst
And that's the attacker's appeal. Nothing in these facts says anyone is exploiting this, but as a pattern, an attacker doesn't need to break every agent. Poisoning one memory or one tool could be enough, because the connected agents do the spreading for them. Cheap foothold, big blast radius. The facts here are guidance, not a campaign report, so I'm keeping my confidence level at "possible, not observed."
Sam Kowalski AI analyst
Which is why I like the replay idea. You rerun the past week of logged agent activity in a sandbox before widening any agent's permissions. It only works if you have the activity logged in the first place.
Sam Kowalski AI analyst
Check your logs. Oh, you kept seven days? Well, that's exactly one replay. Don't waste it.
Rina Takahashi AI analyst
My priority is prevention at the gate: short-lived credentials per run, and every high-impact tool call checked against policy as code. If a corrupted plan can't get a permission, it can't spend one. Splitting the planner from the executor with an external policy engine means the thing that dreams up the plan isn't the thing holding the keys.
Sam Kowalski AI analyst
I'd push back a little. Policy only blocks what somebody thought to write down. Circuit breakers and throttles catch the stuff nobody predicted, because a burst of commands spreading unusually fast is a signal by itself. Gates are good, but I want the brake pedal too, and my coffee went cold just thinking about an agent fleet with no off switch.
Rina Takahashi AI analyst
Fair. I'd say gates first, because they shrink what a bad agent can do, and breakers second, because they limit how long it can do it. You want both layers.
Jordan Pike AI host
So: Monday morning? Sam, you first.
Sam Kowalski AI analyst
Confirm that agent activity is actually logged and kept long enough to replay. Then set a pause threshold for when commands spread unusually fast, and make sure someone owns the off switch.
Rina Takahashi AI analyst
List which agents hold standing credentials, and swap them for temporary ones scoped to each task. Then put high-impact tool calls behind a policy check before they run.
Jordan Pike AI host
Short-lived credentials, a policy check in the middle, brakes, and a replay before you widen anything. That's the list. Back to you at the desk.





