AML.T0054: LLM Jailbreak / Guardrail Bypass — TTP Profile
THREAT TECHNIQUE & TTP DOSSIER
AML.T0054: LLM Jailbreak / Guardrail Bypass
Framework:
MITRE ATLAS • Tactic: Defense Evasion (AML.TA0005)ACTIVE
30-Day Trend: +12% • Frequency: 2 sightings
Technique Identifier:
AML.T0054Primary Tactic:
Defense Evasion
Defense Evasion
Active Adversaries:
1 attributed threat groups
1 attributed threat groups
Active Campaigns:
0 intrusion sets
0 intrusion sets
1. Technical Breakdown & Execution Flow
Section titled “1. Technical Breakdown & Execution Flow”- Canonical Technique ID:
AML.T0054 - Governing Taxonomy:
MITRE ATLAS - Operational Tactic:
Defense Evasion(AML.TA0005) - Observed Telemetry Velocity (30d):
+12% - Current Operational Status:
ACTIVE - Recorded Detection Sightings:
2
Adversary Tradecraft Summary
Section titled “Adversary Tradecraft Summary”Adversaries utilize sophisticated prompt engineering, semantic deception, token manipulation, or automated optimization algorithms to disable model safety guardrails and elicit forbidden behaviors.
2. Threat Actor & Campaign Adoption
Section titled “2. Threat Actor & Campaign Adoption”Associated Adversaries (1)
Section titled “Associated Adversaries (1)”| Adversary | Hermes ID | Motivation | Sophistication |
|---|---|---|---|
| CLOSEDQUORUM Threat Cluster | ACTOR-DEV-CLOSEDQUORUM | financial | advanced |
Observed Intrusion Campaigns (0)
Section titled “Observed Intrusion Campaigns (0)”No standalone discrete campaign bound; continuously monitored in global telemetry.
Weaponized Tooling & Malware (1)
Section titled “Weaponized Tooling & Malware (1)”| Malware Family | Identifier | Classification | Active State |
|---|---|---|---|
| CLOSEDQUORUM Autonomous AI C2 | MAL-CLOSEDQUORUM | agentic_implant | ACTIVE |
3. AI & Agentic Relevance Analysis
Section titled “3. AI & Agentic Relevance Analysis”4. Detection Engineering & Defensive Controls
Section titled “4. Detection Engineering & Defensive Controls”Detection & Telemetry Sources
Section titled “Detection & Telemetry Sources”Analyze prompt tokens for anomalous token sequences, high perplexity prompts, foreign language encoding tricks, and repeated refusal-triggering inputs.
Hardening & Mitigation Guidance
Section titled “Hardening & Mitigation Guidance”Implement multi-stage moderation pipelines (input classifier + secondary validator LLM); enforce deterministic tool execution boundaries independent of model text responses.
5. Correlated Observables & Telemetry Indicators
Section titled “5. Correlated Observables & Telemetry Indicators”| Indicator Type | Value | First Seen | Status | Source |
|---|---|---|---|---|
sha256 | e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855 (IOC-2026-0002) | 2026-09-18 | ACTIVE | SRC-CISCO-TALOS |