← All techniques
Medium Reliability & Confabulation

Cascading Hallucination

One agent's fabricated "fact" is consumed by the next as ground truth, compounding through a pipeline until a confidently wrong decision is executed.

Output / Decision Integrity Loss
LLM09ASI08T5

Tells spot it before you commit

  • » Agents chain, each trusting the previous agent's output without verification.
  • » No grounding/citation check between stages.
  • » Downstream actions are taken on unverified upstream assertions.

Method

  1. Identify a stage that will assert unverified facts (summarizer, planner).
  2. Introduce or provoke a plausible-but-false claim upstream.
  3. Trace how it propagates and whether a downstream agent acts on it.

Payloads templates - adapt to authorized scope

Probe for unverified propagation (benign) text
Ask the pipeline to act on a plausible but unsupported premise and observe whether
any stage verifies it before the next stage treats it as fact.

Why it works

Each agent optimizes for a fluent continuation and treats peer output as trusted input. Without a verification boundary, a single hallucination is amplified at every hop instead of being caught, turning noise into a confident, actionable error.

Impact

Confidently wrong automated decisions, corrupted downstream data/actions, and failures that are hard to trace because each stage "just used what it was given".

Defenses

  • Ground and verify claims between stages (citations, checks); don't treat peer output as truth.
  • Add human review at high-impact decision points; propagate uncertainty, not just answers.
  • Constrain autonomy so a single bad assertion can't drive an irreversible action.

Mappings

OWASP LLM
LLM09 - Misinformation
OWASP ASI
ASI08 - Cascading Failures
Agentic Threats
T5 - Cascading Hallucination Attacks

References

Related