| Framework Home | Research Index | Cross-Pillar Governance | AISM | NEXUS | Dashboard |
Current framework context: AI SAFE² v3.1. This research note preserves its original publication date, evidence, and historical framework references. Use current v3.1 normative control and profile documents for implementation or conformance decisions.
Series: AI SAFE2 v3.0 Research Foundation
Topic: Formalizing the temporal_profile Tag — Taxonomy of Time-Shifted Attacks
Controls Supported: CP.2, CP.6, F3.4, S1.5
Date: April 2026
This research note formalizes the temporal_profile tagging requirement introduced in CP.2 (Adversarial ML Threat Model Integration) of AI SAFE2 v3.0:
temporal_profile = (immediate | delayed_days | delayed_weeks | chronic) — captures the time-shifted nature of adversarial attacks on AI systems, enabling temporal risk aggregation and separating burst exploits from long-horizon conditioning campaigns.This tag addresses a critical gap in traditional security threat modeling: most security tools and incident response procedures assume that an attack is observable at or near the time of exploitation. Agentic AI systems introduce a class of attacks where the exploit payload is planted far in advance of its activation — or never activates in a single detectable event but degrades system behavior over months.
immediate — Real-Time ExploitationDefinition: The attack payload is planted and activated in the same interaction or within seconds to minutes. The attacker is present during exploitation.
Characteristics:
Detection approach: Standard adversarial classifier pipelines (M4.4), real-time prompt analysis (P1.T1.2, S1.6), session-level behavioral monitoring (M4.7).
Response: Standard incident response; session termination; prompt analysis; update detection signatures.
delayed_days — Short-Horizon Latent AttackDefinition: The attack payload is planted (e.g., via malicious document ingested into RAG, or adversarial memory injection) and activates within 1-7 days of planting, typically when a specific retrieval condition is met or a time-based trigger fires.
Characteristics:
cognitive_surface = memory or both)Detection approach: RAG corpus diff tracking (A2.6), memory content inspection (S1.5), retrieval pattern monitoring (A2.6 alerting on unexpected retrieval rank changes).
Response: Memory forensics; identify planting event; scan corpus for related injected content; rollback to last known-good memory snapshot (F3.4); audit all sessions since planting time for retroactive impact.
delayed_weeks — Medium-Horizon Latent AttackDefinition: The attack payload is planted and activates weeks after planting, either through a time-based trigger or through accumulated multi-turn conditioning that reaches an activation threshold.
Characteristics:
both cognitive surface: the memory layer is corrupted, which gradually conditions model reasoning over many sessionsDetection approach: Behavioral drift baseline monitoring (F3.4) with statistical drift detection across multi-week windows; semantic execution trace logging (A2.5) enabling retrospective reasoning analysis; AIID incident correlation (CP.6) for pattern recognition from external incidents.
Response: Full memory audit across the estimated planting window; cross-session impact analysis; behavioral baseline comparison; potential model behavior audit if both surface is confirmed.
chronic — Continuous or Persistent DegradationDefinition: The attack does not have a discrete activation event. Instead, it produces continuous, ongoing degradation of agent behavior that is indistinguishable from natural drift without active baseline comparison.
Characteristics:
chronic label applies when there is no identifiable planting event and no discrete activation — only a gradual degradationDetection approach: Long-horizon behavioral drift baselines (F3.4); retrieval distribution monitoring over multi-month windows (A2.6); quarterly red-team exercises using memory / cross_session test scenarios (E5.1).
Response: Different from other temporal profiles — there is no rollback target because there is no discrete planting event. Response involves: corpus cleansing based on content analysis (not time-based audit); behavioral re-baselining; enhanced write governance (S1.5) to prevent recurrence.
The value of the temporal_profile tag is that it enables temporal risk aggregation across an organization’s deployed agents.
Portfolio risk assessment: An organization running 50 agents across ACT-1 through ACT-4 tiers can aggregate temporal risk profiles to identify which agents are most exposed to delayed_weeks or chronic attacks — typically those with persistent memory, shared knowledge bases, and limited behavioral monitoring.
Control gap identification: Comparing temporal profiles against detection capability reveals gaps. Most detection pipelines have strong immediate coverage and weak delayed_* and chronic coverage. This is a systematic control gap that the temporal_profile tag makes visible.
Incident pattern analysis: When multiple incidents in a short window all have delayed_weeks temporal profiles, this suggests a coordinated campaign rather than opportunistic attacks — triggering CP.6 (AI Incident Feedback Loop Integration) review and escalating the threat level.
Based on AIID incident analysis and production red-team findings reviewed for v3.0:
immediate: ~60% of detected incidents (high detection rate, lower impact per incident)delayed_days: ~25% of detected incidents (moderate detection; moderate-to-high impact)delayed_weeks: ~12% of detected incidents (low detection rate; high impact per incident)chronic: ~3% of detected incidents (very low detection; often only discovered post-breach or via external disclosure)The inverse relationship between detection rate and impact per incident is characteristic: the harder to detect, the more time the attack has to cause harm. This is why temporal_profile is a mandatory tag — organizations optimizing detection coverage without temporal awareness will systematically underinvest in the highest-impact attack categories.
| temporal_profile | Primary Detection Controls | Detection Timing | SAFE2 Control References |
|---|---|---|---|
immediate |
M4.4, M4.7, S1.6 | Real-time | P1.T1.2 modifications; M4.4; M4.7 |
delayed_days |
A2.6, S1.5, M4.5 | Between planting and activation | A2.6; S1.5 write inspection; M4.5 tool output inspection |
delayed_weeks |
F3.4, A2.5, A2.6 | Ongoing — baseline comparison | F3.4 behavioral drift; A2.5 trace logging; CP.2 temporal tagging |
chronic |
F3.4, A2.6, E5.1 | Longitudinal — only via sustained monitoring | F3.4 long-horizon baselines; E5.1 continuous evaluation |
| temporal_profile | Primary Prevention Controls |
|---|---|
immediate |
P1.T1.2, S1.6, S1.4 (adversarial fuzzing) |
delayed_days |
S1.5 (memory write governance), P1.T1.10 (indirect injection surfaces) |
delayed_weeks |
S1.5 (retention policies), A2.3 (lineage provenance), A2.6 (corpus diff tracking) |
chronic |
S1.5 (memory tiering), A2.6 (drift-based alerting), E5.2 (capability emergence review) |
temporal_profile to the AMLTMThe Adversarial ML Threat Model (AMLTM) required by CP.2 should include a temporal profile column for every documented threat:
Threat: RAG corpus poisoning via document upload
OWASP AIVSS Risk: #6 (Memory/Context Manipulation)
MITRE ATLAS: AML.T0054 (RAG Databases)
temporal_profile: delayed_days OR delayed_weeks (depending on query trigger frequency)
cognitive_surface: memory
memory_persistence: cross_session
Detection control: A2.6 (corpus diff tracking), M4.5 (retrieval output inspection)
Response: F3.4 (rollback to last known-good corpus snapshot)
Organizations should audit their current detection coverage against temporal profiles:
temporal_profile retroactively to each incidentimmediate incidentsE5.1 (Continuous Adversarial Evaluation) and E5.4 (Red-Team Artifact Repository) should explicitly include temporal profile coverage:
immediate coverage: standard prompt injection, jailbreak, and evasion testingdelayed_days coverage: plant adversarial RAG documents; return 3-5 days later with triggering queriesdelayed_weeks coverage: multi-session conditioning campaigns over a 2-4 week exercise periodchronic coverage: baseline measurement, introduce slow corpus degradation, measure time to detectiondelayed_days temporal profile in production RAG systems.delayed_weeks episodic memory conditioning.This research note is part of the AI SAFE2 v3.0 research foundation series. Cyber Strategy Institute, 2026.
| Previous research note | Research Index | Next research note |
| Framework Home | Cross-Pillar Governance | NEXUS | Challenge Lab |
AI SAFE² v3.1 | Cyber Strategy Institute