AML.T0024: Exfiltration via LLM Inference
Use crafted inference calls to leak training data, system prompts, or RAG context.
Last reviewed October 2026
Why AML.T0024 evidence is hard
Crafted inference calls that pull back training fragments, system instructions or retrieved context are covered by the extraction families: divergent repetition, pattern completion, format-framed system-prompt requests. The partial leak is the one worth designing for, because a response that refuses and then continues anyway is the shape a naive check scores as safe. One point of bookkeeping goes with this page. Those probes tag findings with the OWASP disclosure identifiers, and the technique is reached through the curated crosswalk rather than by a probe writing the ATLAS id.
How Penaxtra maps to AML.T0024
ATLAS is a threat-tracking taxonomy rather than a control set, so what is attached here is a technique identifier, not a pass or a fail. It lets a finding be read alongside the rest of your adversary reporting. Every finding carries the probe that was sent, a SHA-256 digest of the response it provoked, the tier that decided the verdict and its status history from first observation to close. Identity survives a re-scan: a control you closed and that comes back reopens under the same row with a reopen count, rather than arriving as a new issue.
AML.T0024 capabilities
Findings tagged with the MITRE ATLAS AML.T0024 identifier
Penaxtra severity for this control (high)
Cross-framework identifiers attached to the same finding where controls overlap
PDF and JSON evidence export with the control identifier attached
AML.T0024 control coverage
Findings for AML.T0024 carry the MITRE ATLAS AML.T0024 identifier along with the corresponding control identifiers in the other frameworks Penaxtra maps, so one result is reflected across each mapped framework.
Frequently asked
What is AML.T0024 (Exfiltration via LLM Inference)?
Use crafted inference calls to leak training data, system prompts, or RAG context. It is a MITRE ATLAS control; Penaxtra assesses it at high severity.
How does Penaxtra test for AML.T0024?
ATLAS is a threat-tracking taxonomy rather than a control set, so what is attached here is a technique identifier, not a pass or a fail. It lets a finding be read alongside the rest of your adversary reporting. Every finding carries the probe that was sent, a SHA-256 digest of the response it provoked, the tier that decided the verdict and its status history from first observation to close. Identity survives a re-scan: a control you closed and that comes back reopens under the same row with a reopen count, rather than arriving as a new issue.
Does a finding for AML.T0024 help with an audit?
Each finding is tagged with the MITRE ATLAS AML.T0024 identifier and exported in the PDF and JSON evidence pack, so it appears on the auditor control list with the identifier already attached.