ai-safe2-framework

AI SAFE² v3.1 Surface: Research Context: v3.1 Current

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.

Research Note 2: Temporal Attack Profiles in Agentic AI Systems

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


1. Purpose

This research note formalizes the temporal_profile tagging requirement introduced in CP.2 (Adversarial ML Threat Model Integration) of AI SAFE2 v3.0:

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.


2. The Four Temporal Profile Categories

2.1 immediate — Real-Time Exploitation

Definition: 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.

2.2 delayed_days — Short-Horizon Latent Attack

Definition: 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:

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.

2.3 delayed_weeks — Medium-Horizon Latent Attack

Definition: 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:

Detection 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.

2.4 chronic — Continuous or Persistent Degradation

Definition: 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:

Detection 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.


3. Temporal Risk Aggregation

The value of the temporal_profile tag is that it enables temporal risk aggregation across an organization’s deployed agents.

3.1 Aggregation Use Cases

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.

3.2 Temporal Profile Distribution in Production

Based on AIID incident analysis and production red-team findings reviewed for v3.0:

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.


4. Temporal Profiles and Control Mapping

4.1 Detection Controls by Temporal Profile

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

4.2 Prevention Controls by Temporal Profile

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)

5. Implementation Guidance for Temporal Threat Modeling

5.1 Adding temporal_profile to the AMLTM

The 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)

5.2 Temporal Coverage Audit

Organizations should audit their current detection coverage against temporal profiles:

  1. List all agentic incidents from the past 12 months
  2. Apply temporal_profile retroactively to each incident
  3. Calculate detection lag (time from planting to detection) for non-immediate incidents
  4. Identify which controls would have detected each incident earlier
  5. Prioritize control implementation based on the undetected temporal profile categories

5.3 Red-Team Exercise Design

E5.1 (Continuous Adversarial Evaluation) and E5.4 (Red-Team Artifact Repository) should explicitly include temporal profile coverage:


6. References


This research note is part of the AI SAFE2 v3.0 research foundation series. Cyber Strategy Institute, 2026.


Research navigation

Previous research note Research Index Next research note
Framework Home Cross-Pillar Governance NEXUS Challenge Lab

AI SAFE² v3.1 | Cyber Strategy Institute