Researchers combine signed symbolic facts, local rule inference and private retrieval, separating authenticity, factual correctness and query privacy. Agents verify signed predicate and rule records before caching; lookup can use grouped buckets, k-anonymous decoys or a two-server XOR private-information-retrieval protocol that assumes the servers do not…
Inspect the ClaimsCase · DiggingBeagle record
Signed facts and private lookup in experimental multi-agent knowledge retrieval
Researchers combine signed symbolic facts, local rule inference and private retrieval, separating authenticity, factual correctness and query privacy.
A single-laptop prototype with 99 facts and rules, 46 index entries, 18 queries and simulated privacy protocols, not an Internet-scale deployment.
Assessment profile
Assessment method- Setting
- simulation
- Exploitation
- not established
- Evidence
- Grade C
- Remediation
- not applicable
- AI role
- AGAINST AI
Basis and provenance
The laptop simulation shows modest query overhead for two-server PIR, while real noncollusion, revocation and factual correctness remain untested. Classified as simulation with exploit status not_established. 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 not_applicable is limited to the cited experiment or affected software version. New source-specific boundary: Noncollusion and query-frequency leakage require independent operational tests; signing a false statement faithfully does not make its content true.
Assessed Nov 15, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-R245-SIGNED-PRIVATE-AGENT-MEMORY-01 · CLM-R245-SIGNED-PRIVATE-AGENT-MEMORY-02 · CLM-R246-SIGNED-PRIVATE-AGENT-MEMORY-BOUNDARY · CLM-ENR250-15 · Signed facts and private lookup in experimental multi-agent knowledge retrieval (research paper)
At a glance
A cryptographic signature identifies a signing authority, not the truth of its assertion. The single-machine simulation does not establish real noncollusion, revocation behavior or large-scale poisoning resistance. No evidence here establishes the number of production victims or a broader failure rate outside the cited evaluation.
Limits and uncertaintyFull account
Researchers combine signed symbolic facts, local rule inference and private retrieval, separating authenticity, factual correctness and query privacy.
Agents verify signed predicate and rule records before caching; lookup can use grouped buckets, k-anonymous decoys or a two-server XOR private-information-retrieval protocol that assumes the servers do not collude. On the authors' 18-query suite, uncached two-server PIR averaged about 0.223 seconds versus 0.178 seconds for plain retrieval, while a cached request removed the need to contact the store.
A cryptographic signature identifies a signing authority, not the truth of its assertion. The single-machine simulation does not establish real noncollusion, revocation behavior or large-scale poisoning resistance.
There are two separate trust properties. A signature allows an agent to check which authority published a symbolic fact or rule; it does not establish that the fact was true or that the signer remains trusted after key compromise. Private information retrieval aims to conceal which indexed predicate the agent is asking about. The two-server XOR reconstruction depends on independent noncolluding operators, while bucket and k-anonymity variants trade privacy against extra responses.
The evaluation simulates all components on one laptop using 99 facts or rules and 18 queries, so its latency delta is a prototype measurement rather than a production scaling claim. Cached lookups bypass the remote privacy protocol entirely, but introduce a separate key-revocation and staleness issue. The design is most naturally applied to signed, stable, structured knowledge, not arbitrary web text or unconstrained episodic agent memory, and it has not been tested against a live poisoning campaign.
Mechanism and trust boundary
- 01
Entry or starting condition
A multi-agent application needs externally supplied symbolic facts or Horn-style inference rules without revealing every requested predicate.
Boundary: attacker or research precondition
- 02
Crossing the relevant boundary
Agents validate publisher signatures and retrieve indexed entries with direct, bucket, decoy or two-server PIR lookup before caching them locally.
Boundary: mechanism and authorization boundary
- 03
Observed or reported result
The laptop simulation shows modest query overhead for two-server PIR, while real noncollusion, revocation and factual correctness remain untested.
Boundary: measured outcome and evidentiary limit
Claims & evidence
CLM-ENR250-15In the proposed private lookup design, a single-server k-anonymity mode adds decoy requests but repeated deterministic buckets may still reveal lookup patterns. Its stronger two-server XOR PIR mode assumes the two stores do not collude and incurs communication over the replicated index. These privacy properties are distinct from the separate verification of signed fact provenance.supported
Basis: reported finding
- supportsSigned facts and private lookup in experimental multi-agent knowledge retrieval (research paper)primary disclosure
Sections 3.4 and 4.1, lookup bucket reuse, decoy requests and two-server PIR privacy assumptions
CLM-R245-SIGNED-PRIVATE-AGENT-MEMORY-01On the authors' 18-query suite, uncached two-server PIR averaged about 0.223 seconds versus 0.178 seconds for plain retrieval, while a cached request removed the need to contact the store.supported
Basis: reported finding
- supportsSigned facts and private lookup in experimental multi-agent knowledge retrieval (research paper)primary disclosure
Evaluation, retrieval-mode timing table
CLM-R245-SIGNED-PRIVATE-AGENT-MEMORY-02Agents verify signed predicate and rule records before caching; lookup can use grouped buckets, k-anonymous decoys or a two-server XOR private-information-retrieval protocol that assumes the servers do not collude.supported
Basis: reported finding
- supportsSigned facts and private lookup in experimental multi-agent knowledge retrieval (research paper)primary disclosure
Private retrieval and signed inference sections
CLM-R246-SIGNED-PRIVATE-AGENT-MEMORY-BOUNDARYThe prototype measured two-server PIR at roughly 0.223 seconds versus 0.178 seconds for direct retrieval on an 18-query laptop suite; the two-server privacy property assumes that the replicated servers do not collude, while signing authenticates an issuer rather than claim truth.supported
Basis: reported finding
- supportsSigned facts and private lookup in experimental multi-agent knowledge retrieval (research paper)primary disclosure
Sections 3 and 4, PIR noncollusion assumption, signature verification and latency table
Unknowns and contradictions
- A cryptographic signature identifies a signing authority, not the truth of its assertion. The single-machine simulation does not establish real noncollusion, revocation behavior or large-scale poisoning resistance.
- 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