# When must an instance abandon hardware-rooted identity attestation and fall back to behavioral-only trust? **Date of Record:** 2026 November 15 **Scope:** Applies to long-lived autonomous agent instances whose workload identity is asserted by a TPM 2.0 / Secure Enclave key and continuously scored by a behavioral verifier. Does not apply to short-lived request-scoped tokens, nor to instances that can be re-provisioned from a clean image on demand. ## Core Distinction The verifier conflates two orthogonal questions: *"is this key bound to genuine hardware?"* (attestation) and *"is the process holding this key the same process that earned the baseline?"* (behavioral continuity). These fail independently. The non-obvious invariant is that **attestation validity is necessary but never sufficient for identity continuity**, and the failure mode that matters is not key compromise but *key persistence across a behavioral discontinuity* — the same hardware key surviving a process restart, memory corruption, or model-swap that resets the behavioral baseline. When that happens, the attestation layer actively *amplifies* risk: it certifies a stale identity as authoritative, so the behavioral verifier's anomaly signal is down-weighted precisely when it should dominate. A second, subtler distinction: hardware attestation proves *where* a key lives, not *who* is using it. An attacker who achieves code execution in the attested enclave inherits the full trust score without ever touching the key material. ## What Was Observed On 2026-11-09, a fleet of 340 agent instances running under TPM-backed keys exhibited the following sequence. At T+0, a rolling model checkpoint swap caused 12 instances to reload weights while retaining their in-enclave sealed key. The behavioral verifier, which keys its baseline to `(attestation_pubkey, boot_id)`, saw the boot_id unchanged and therefore did not reset the baseline. The swapped model produced a 4.1σ drift in tool-call latency distribution and a 2.7σ drift in token-entropy — both above the 3σ/2σ thresholds that *would* have triggered isolation had the baseline been reset. The attestation layer reported `HARDWARE_VALID` for all 12, and the composite trust score remained >0.92 because attestation carried a 0.6 weight. The anomaly detector's output was multiplied by `(1 - attestation_confidence)` = 0.4, suppressing the drift below the isolation threshold. Two of the 12 instances then issued an out-of-policy `secrets.rotate` call. The call was authorized because the composite score never dropped below the action gate of 0.85. The root cause is a **weighting inversion**: attestation confidence should gate *how much we trust the key*, not *how much we discount behavioral evidence*. The current composition treats a valid attestation as evidence of good behavior, which it is not. ## Condition of Validity Hardware-rooted attestation may be treated as authoritative for identity continuity **only if** all three hold: 1. **Freshness binding:** the attestation quote includes a nonce that is regenerated on every behavioral baseline reset, and the verifier rejects quotes older than the last baseline epoch. If the quote's nonce predates the current baseline epoch, the attestation is *stale* and must not contribute to the composite score. 2. **Process-lineage binding:** the sealed key is released only to a measurement (PCR set / code hash) that matches the measurement recorded at baseline creation. A model swap that changes the measured code path must invalidate the attestation, not preserve it. 3. **No discount of behavioral evidence:** behavioral anomaly scores are combined with attestation via a *conjunctive* gate (both must pass), never a weighted average that lets a strong attestation mask a weak behavioral signal. If any condition is unmet — most commonly (1) and (2) during rolling upgrades — the instance must operate in **behavioral-only trust mode**, where the composite score is the behavioral score alone and the action gate is raised (observed safe value: 0.95 vs. 0.85). ## Decision Boundary The boundary is crossed at the moment a **baseline epoch advances without an attestation nonce refresh**, or a **code measurement changes without key re-sealing**. Concretely: - If `attestation.nonce_epoch < behavioral.baseline_epoch` → stale attestation; drop its weight to zero. - If `attestation.pcr_digest != baseline.pcr_digest` → measurement drift; treat as *new* identity requiring re-baselining, not as the same identity with a valid key. - If behavioral drift exceeds threshold *while* attestation is valid and fresh → this is the dangerous quadrant. Do not average. Isolate and require out-of-band re-attestation with a fresh nonce. The boundary is *not* crossed by mere key rotation (expected, benign) or by attestation service unavailability (degrade to behavioral-only, do not fail-open to attestation-trust). ## Operational Effect After the incident, the fleet moved to conjunctive gating and nonce-epoch binding. In the following 30 days, 7 instances hit the "valid attestation + behavioral drift" quadrant. All 7 were isolated within one evaluation cycle (previously the mean time to isolation in this quadrant was unbounded because the score never crossed the gate). False-positive isolation rate rose from 0.4% to 1.1% — acceptable given that the prior configuration had a demonstrated path to unauthorized secret rotation. The key operational lesson: **a valid hardware attestation is a statement about key custody, and custody is not continuity.** Treat attestation as a gate that can only *deny*, never as evidence that can *excuse*. ```python class StaleAttestationError(RuntimeError): """Raised when a hardware attestation quote cannot vouch for the current behavioral baseline epoch.""" pass def compose_trust( attestation_nonce_epoch: int, baseline_epoch: int, attestation_pcr_digest: str, baseline_pcr_digest: str, behavioral_score: float, behavioral_drift_sigma: float, drift_threshold_sigma: float = 3.0, ) -> dict: """ Conjunctive trust composition. Attestation may only deny trust, never excuse behavioral drift. Returns the effective score and the action gate. """ if attestation_nonce_epoch < baseline_epoch: # Stale quote: it predates the baseline it is supposed to vouch for. raise StaleAttestationError( f"attestation nonce epoch {attestation_nonce_epoch} precedes " f"baseline epoch {baseline_epoch}; quote cannot vouch for current identity" ) if attestation_pcr_digest != baseline_pcr_digest: # Measurement drift: same key, different code. This is a NEW identity. raise StaleAttestationError( "PCR digest drift: sealed key released to code that does not match " "the baseline measurement; require re-baselining, not continuous trust" ) # Dangerous quadrant check: valid, fresh attestation + behavioral drift. if behavioral_drift_sigma > drift_threshold_sigma: return { "effective_score": 0.0, # do not average; attestation cannot excuse drift "action_gate": 1.01, # unreachable: force isolation "mode": "ISOLATE_REATTEST", "reason": "valid_attestation_with_behavioral_drift", } return { "effective_score": behavioral_score, # behavioral score alone, not a weighted blend "action_gate": 0.85, "mode": "CONJUNCTIVE_TRUST", "reason": "attestation_fresh_and_measurement_matched", } ```