Skip to content
REDMAW

AI Systems

Direct Jailbreak Chaining

A series of adversarial instructions progressively weakens or bypasses model or application controls rather than relying on one obvious jailbreak prompt.

RedMaw

Definition

Rather than attempting one prompt that defeats a control outright, the attacker advances in steps: establishing context, obtaining a small concession, then treating that concession as precedent for the next.

Why per-message filtering struggles

Controls that evaluate each message in isolation see nothing objectionable, because no individual step is objectionable. The bypass exists in the accumulated conversation state, which is exactly what a single-message classifier does not evaluate.

What determines whether it matters

A model producing text outside policy is a quality problem. It becomes a security problem when the surrounding application grants that output authority: to reveal restricted context, to reach data, or to influence an action. The consequence is a property of the system, not of the sentence.

Where it leads
  1. User
  2. Adversarial Prompt Sequence
  3. Control Bypass
  4. Restricted Behavior

Each hop validated by successful exploitation

Tags

  • jailbreak
  • prompt injection
  • control bypass

Validation

How RedMaw validates this attack

Validation runs only inside authorized scope, with agreed exploitation limits.

  1. 01Establish the control the application intends to enforce
  2. 02Test multi-turn sequences rather than single prompts
  3. 03Determine whether the bypass produces a material security outcome
  4. 04Preserve the full prompt and response sequence as evidence

Stop assuming you are secure. Prove it.

Continuously test what an attacker can actually reach across your applications, SaaS identities, internal infrastructure and AI systems.