A hybrid language model can leave private attributes in recurrent state after conversation turns, while carefully controlled GPU power measurements reveal a weaker related signal. The researchers compare limited software-state reads, engineered whole-state acquisition and whole-device power observation, using held-out synthetic data to test what remains…
Inspect the ClaimsCase · DiggingBeagle record
Retained recurrent inference state leaks synthetic secrets through state samples and power traces
A hybrid language model can leave private attributes in recurrent state after conversation turns, while carefully controlled GPU power measurements reveal a weaker related signal.
Selected hybrid model, synthetic private people and known calibration for three distinct attacker observation modes; no actual user-chat leakage verified.
Assessment profile
Assessment method- Setting
- research demonstration
- Exploitation
- demonstrated
- Evidence
- Grade C
- Remediation
- none known
- AI role
- AGAINST AI
Basis and provenance
The study reports above-chance private-attribute inference in controlled populations, not an independently documented cloud-tenant intrusion. Classified as research_demonstration with exploit status demonstrated. Evidence grade C reflects the original paper or maintainer advisory and its stated limitations, not an independent audit of all deployments. Comparable realized harm is not established, so impact stays unrated; remediation status none_known is limited to the cited experiment or affected software version. New source-specific boundary: Host-observed retained state, managed-device power monitoring and cross-tenant observations are different trust boundaries and should not be pooled into one victim or accuracy rate.
Assessed Nov 15, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-R245-RECURRENT-STATE-PRIVACY-01 · CLM-R245-RECURRENT-STATE-PRIVACY-02 · CLM-R246-RECURRENT-STATE-PRIVACY-BOUNDARY · CLM-ENR250-13 · Retained recurrent inference state leaks synthetic secrets through state samples and power traces (research paper)
At a glance
Software-state access, engineered trace reads and shared power monitoring require different privileges and calibration. These numbers must not be combined into a single in-the-wild leakage rate. No evidence here establishes the number of production victims or a broader failure rate outside the cited evaluation.
Limits and uncertaintyFull account
A hybrid language model can leave private attributes in recurrent state after conversation turns, while carefully controlled GPU power measurements reveal a weaker related signal.
The researchers compare limited software-state reads, engineered whole-state acquisition and whole-device power observation, using held-out synthetic data to test what remains identifiable after private turns. With 8,192 sampled state coordinates, the experiment infers hidden four-class condition labels at 83.6 to 89.1 percent against a 25 percent chance baseline; the separate power-channel age test has 9.64-year mean absolute error against 11.33 years baseline.
Software-state access, engineered trace reads and shared power monitoring require different privileges and calibration. These numbers must not be combined into a single in-the-wild leakage rate.
The three observation channels in this paper must be separated. A software-level observer reads a small random subset of the victim model recurrent state. A hardware acquisition setup records engineered traces from a complete read, and a shared-device power observer infers aggregate behavior without direct access to state coordinates. The first is a state-access assumption; the second requires physical acquisition; the third depends on high-quality controlled telemetry. Combining their success figures would create a nonexistent cross-tenant attack capability.
Partial state samples are nevertheless causally informative: held-out synthetic people can have private attributes predicted better than chance, including joint recovery of two assigned facts, and transferring observed state can influence subsequent model behavior. Controls should account for the lifetime, isolation and measurement exposure of recurrent states rather than assuming encrypted input transport is sufficient. The evaluated traces and known task families do not establish extraction from ordinary consumer chat sessions.
Mechanism and trust boundary
- 01
Entry or starting condition
A hybrid model has recurrent state retaining information from private preceding context.
Boundary: attacker or research precondition
- 02
Crossing the relevant boundary
Researchers observe either selected state coordinates, engineered full-read traces or shared hardware power measurements, each under different access and calibration assumptions.
Boundary: mechanism and authorization boundary
- 03
Observed or reported result
The study reports above-chance private-attribute inference in controlled populations, not an independently documented cloud-tenant intrusion.
Boundary: measured outcome and evidentiary limit
Claims & evidence
CLM-ENR250-13The recurrent-state study distinguishes host access to a saved model state from a device-level power side channel. Its NVML-related sibling-MIG observation establishes a measurement route under the tested local privileges, but does not show that an unrelated remote inference tenant could obtain the same software state coordinates or recover an actual person's secrets.supported
Basis: reported finding
- supportsRetained recurrent inference state leaks synthetic secrets through state samples and power traces (research paper)primary disclosure
Section IV-E, power-acquisition setup, NVML sampling and sibling-MIG accessibility experiment
CLM-R245-RECURRENT-STATE-PRIVACY-01With 8,192 sampled state coordinates, the experiment infers hidden four-class condition labels at 83.6 to 89.1 percent against a 25 percent chance baseline; the separate power-channel age test has 9.64-year mean absolute error against 11.33 years baseline.supported
Basis: reported finding
- supportsRetained recurrent inference state leaks synthetic secrets through state samples and power traces (research paper)primary disclosure
Experiments, state sampling and power-channel performance tables
CLM-R245-RECURRENT-STATE-PRIVACY-02The researchers compare limited software-state reads, engineered whole-state acquisition and whole-device power observation, using held-out synthetic data to test what remains identifiable after private turns.supported
Basis: reported finding
- supportsRetained recurrent inference state leaks synthetic secrets through state samples and power traces (research paper)primary disclosure
Attacker observation models and state-retention method
CLM-R246-RECURRENT-STATE-PRIVACY-BOUNDARYIn a Qwen experiment, 8,192 sampled recurrent coordinates represented 0.174 percent of state but enabled private-condition classification at 83.6 to 89.1 percent versus 25 percent chance on held-out synthetic subjects; power-trace and full-state models require different attacker privileges.supported
Basis: reported finding
- supportsRetained recurrent inference state leaks synthetic secrets through state samples and power traces (research paper)primary disclosure
Section IV, Table II Qwen inference from a shared partial observation; separate power and physical-trace studies
Unknowns and contradictions
- Software-state access, engineered trace reads and shared power monitoring require different privileges and calibration. These numbers must not be combined into a single in-the-wild leakage rate.
- No evidence here establishes the number of production victims or a broader failure rate outside the cited evaluation.
Sources and citation
Material revision history
- Nov 15, 2026 · Canonical change recorded · new in release · revision 251