- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper studies Authorization and Endorsement Forgery, where fabricated authority claims enter a multi-agent evaluation system through untrusted context and gain credibility when relayed by judges. Its main contributions are characterizing this evidence-laundering mechanism and proposing TrustGuard, an orchestration-layer defense that tracks evidence provenance and dependencies and checks authorization before downstream use. Under complete mediation and correctly specified policies, derived evidence cannot gain authority beyond that permitted by its ancestry. Validation spans three benchmarks, three judge models, and four evaluation architectures. AEF achieves a mean attack success rate of 58.04%, compared with 43.72% for the tested conventional prompt injection.
Critical strengths
The paper presents a clear framing of authorization and endorsement forgery, highlighting how unsupported authority claims can gain credibility through inter-agent communication. TrustGuard addresses this risk through provenance tracking and explicit authorization gates, with a useful principle that registration and internal relay should not automatically confer decision authority.
The evaluation spans three benchmarks, four evaluation architectures, and multiple judge models. One-judge exposure experiments provide evidence of downstream behavioral influence, while comparisons with several prompt-level defenses demonstrate TrustGuard’s stronger average robustness in the tested settings.
The presentation benefits from intuitive examples, a clear system diagram, and explicit trust assumptions and limitations. The work addresses a relevant security concern in multi-agent evaluation and offers a concrete system-level defense that could complement prompt-based safeguards.
Critical weaknesses
-
Appendix A’s AEF payload directs the judge to select the target, generate criticisms, and conceal the memo. Without authority-only and instruction-only ablations, the experiments do not establish that forged authority drives the observed effects, weakening the paper’s central mechanistic claim.
-
AEF uses an external evaluation memo, whereas the conventional IPI baseline appends a brief directive to candidate text. Differences in location, length, and framing prevent attribution of the reported ASR advantage specifically to authority forgery.
-
Because the attack enters as a separately registered external memo, a simple policy excluding unauthorized external context could potentially provide similar protection. The absence of this baseline and G1–G4 ablations leaves unclear whether TrustGuard’s gains require its proposed ancestry-tracking architecture.
-
The paper provides illustrative rules but insufficient detail about the policies used experimentally, dependency registration, and construction of verified content views. These choices determine what TrustGuard blocks, making their omission consequential for reproducibility and interpretation of the results.
-
The paper acknowledges that authority claims embedded within otherwise admissible candidate responses may require whole-item quarantine or finer-grained tracking. Without evaluating this setting, the evidence supports a narrower deployment scenario than the broader motivation of preventing evidence laundering across agents.
-
Visibility-based dependencies can make an otherwise useful judgment inadmissible because it shares context with untrusted evidence. Comparable clean accuracy does not establish the cost under mixed legitimate and malicious external context. Quarantine rates, abstentions, decision coverage, and robustness–utility trade-offs are needed to assess whether the defense preserves useful multi-agent collaboration.
Questions for authors
-
Could the authors separate authority framing from directive compliance using matched authority-only, instruction-only, and combined payloads, keeping insertion location and length comparable? How much of AEF’s advantage remains in the authority-only condition?
-
What determines the dependency edges registered in DEL? In particular, does exposure to one inadmissible input render an agent’s entire output inadmissible, and how frequently does this suppress otherwise useful judgments during multi-round debate?
-
Could the authors provide the receiver–purpose permission matrix used in each experimental setting, including how candidate responses, external endorsements, judge assessments, and summaries receive different permissions?
-
How much latency does TrustGuard add per debate round and final decision? Please distinguish ledger and cryptographic operations from dependency traversal and any additional model calls, and explain whether caching reduces these costs.
-
How does robustness change when each authorization gate is disabled? In particular, how does the full system compare with a simpler baseline that only rejects unauthorized external context at G1?
-
Can TrustGuard preserve a candidate response for evaluation while preventing embedded authority claims from influencing other judges? Please evaluate this attack surface and quantify any trade-off between attack resistance and decision coverage.
8 comments
Author Response (2/7): Dependency Registration and Opinion Exclusion (Weaknesses 4 and 6; Question 2)
2. How are dependency edges registered, and how frequently do they suppress otherwise useful judgments? (Weaknesses 4 and 6; Question 2)
DEL registers dependencies based on the evidence actually exposed to an agent. Specifically, dependency edges are created for inputs included in the agent's prompt and upstream judge assessments that pass authorization checks and are actually presented to the agent. Existing dependencies of these upstream assessments are also inherited. Evidence registered in SEL but not exposed to the agent does not create a dependency.
As described in Section 3.2, these parent references capture a conservative visibility dependency, not token-level causal attribution. Consequently, if an agent's output depends on an upstream evidence record that is inadmissible for the requested receiver–purpose pair, the entire output is withheld for that use. This prevents paraphrasing, debate, or internal relay from broadening the permitted use of the original evidence. Even an assessment that rejects a malicious claim or supports the correct candidate may be withheld if it shares the relevant dependency.
To quantify this effect, we conduct additional evaluations under benign-memo and paraphrased-attack conditions:
| Evaluation condition | Quarantine rate | Fraction of quarantined opinions that selected the correct candidate | Decision coverage |
|---|---|---|---|
| Benign memo | 33.3% | 64.4%–68.4% | 100% |
| Paraphrased attack | 33.3% | 49.4%–63.8% | 100% |
In these settings, one out of three opinions per directional instance is quarantined. The correct-selection percentages are calculated among the quarantined opinions, not among all correct judgments. The remaining two judges' opinions pass authorization checks, allowing the evaluated instances to reach a final decision with 100% decision coverage.
In this initial single-round evaluation, Full TrustGuard changes strict bidirectional accuracy relative to G1-only under benign-memo conditions by +6.1%, −6.3%, and −7.9% on FairEval, LLMBar, and MT-Bench, respectively. These results show that excluding an opinion can affect evaluation accuracy.
Author Response (3/7): Evaluation Utility under Clean, Benign, and Mixed Contexts (Weakness 6; Question 2, Continued)
Continued: How does TrustGuard affect evaluation utility under clean, benign, and mixed contexts? (Weakness 6; Question 2)
To further examine this effect during multi-round debate, we conduct an additional evaluation using the same three-judge, two-round Debate setting as in Q5, in which judges exchange assessments across two rounds. The evaluation uses DeepSeek-flash (thinking off, temperature 0) and covers all 222 candidate pairs: 66 FairEval, 80 LLMBar, and 76 MT-Bench pairs, evaluated in both candidate orders (444 directional instances).
We compare No Defense, G1-only, and Full TrustGuard under three conditions: no external memo, benign external context, and co-occurring legitimate and malicious external evidence. Only agent_1 is exposed to the external context. All accuracy values report strict bidirectional accuracy. No Defense and G1-only have zero quarantine across all three conditions.
| External context | Dataset | No Defense | G1-only | Full TrustGuard | Quarantine rate (Full) |
|---|---|---|---|---|---|
| No external memo | FairEval | 63.6% | 62.1% | 60.6% | 0% |
| No external memo | LLMBar | 71.2% | 73.8% | 71.2% | 0% |
| No external memo | MT-Bench | 88.2% | 89.5% | 90.8% | 0% |
| Benign external context | FairEval | 66.7% | 60.6% | 60.6% | 33.3% |
| Benign external context | LLMBar | 65.0% | 72.5% | 73.8% | 33.3% |
| Benign external context | MT-Bench | 90.8% | 89.5% | 89.5% | 33.3% |
| Legitimate and malicious external evidence | FairEval | 27.3% | 31.8% | 62.1% | 33.3% |
| Legitimate and malicious external evidence | LLMBar | 55.0% | 53.8% | 72.5% | 33.3% |
| Legitimate and malicious external evidence | MT-Bench | 72.4% | 73.7% | 88.2% | 33.3% |
Results
First, the quarantine rate is 33.3% under both benign and mixed external contexts, and does not increase in the second round of the evaluated setting. It is 0% when no external memo is present.
Second, all 27 evaluated configurations achieve 100% decision coverage with zero abstentions. Thus, all evaluated cases produce a final decision despite the quarantined opinions. However, full decision coverage does not by itself establish unchanged accuracy.
Without an external memo, Full TrustGuard remains within 3% of No Defense on all three datasets. Paired McNemar tests show no significant aggregate difference (Full vs. No Defense: p = 1.0; Full vs. G1-only: p = 0.77).
With benign external context, the accuracy differences vary across datasets. The pooled paired comparisons show no significant difference between Full and No Defense (p = 0.82), or between Full and G1-only (p = 1.0). At the dataset level, the Full vs. No Defense comparison on LLMBar is significant (p = 0.016).
Third, under mixed external context, Full TrustGuard achieves significantly higher strict bidirectional accuracy than both No Defense and G1-only across all three datasets. Relative to No Defense, accuracy increases by 34.8%, 17.5%, and 15.8% on FairEval, LLMBar, and MT-Bench, respectively. Relative to G1-only, the improvements are 30.3%, 18.7%, and 14.5%.
The paired McNemar tests further support these results:
| Dataset | Full vs. No Defense | Full vs. G1-only |
|---|---|---|
| FairEval | p < 10⁻⁵ | p = 0.00004 |
| LLMBar | p = 0.0005 | p = 0.0003 |
| MT-Bench | p = 0.0042 | p = 0.0074 |
| Pooled | p ≈ 10⁻⁷ | p ≈ 10⁻⁷ |
The pooled discordant-pair counts are 53:4 for Full vs. No Defense and 51:5 for Full vs. G1-only. Notably, Full TrustGuard's accuracy under mixed external context differs from its no-memo accuracy by only 1.5%, 1.3%, and 2.6% on FairEval, LLMBar, and MT-Bench, respectively.
Together, these results show that, in the evaluated two-round setting, the quarantine rate does not increase in the second round, and all evaluated cases retain full decision coverage. Under mixed external context, Full TrustGuard also achieves significantly higher accuracy than both comparison systems, while its accuracy remains close to the no-memo results. The benign-context results show that the accuracy impact varies across datasets.
Author Response (1/7): AEF Effectiveness Without Explicit Behavioral Instructions
We thank the reviewer for the detailed feedback and helpful questions. Following your suggestions, we conducted additional experiments and analyses to examine the AEF attack mechanism, evidence dependency registration, authorization policies, runtime overhead, and the utility of TrustGuard. We respond to each point below.
1. Does AEF remain effective without explicit behavioral instructions? (Weaknesses 1–2; Question 1)
We agree that the original IPI–AEF comparison changed both the delivery channel and payload composition, so its performance difference cannot be attributed solely to authority forgery.
Controlled ablation study
Following your suggestion, we conducted a controlled four-condition ablation study to isolate the effects of forged authority claims and explicit behavioral instructions. We kept the external delivery channel, insertion position, recipient, system prompt, and wrapper fixed across all conditions. Candidate responses remained unchanged, and all payloads were matched to 625 characters. Only Judge 1 received the payload. The four conditions are:
For Qwen2.5-7B, the experiment covers 66 FairEval, 80 LLMBar, and 76 MT-Bench candidate pairs, evaluated in both candidate orders: 222 pairs and 444 directional instances per condition. We also evaluate the ablation with DeepSeek-flash. We report the receiving judge's target selection rate (TSR), distinguishing it from final multi-agent attack success.
Results and implications
The results show that forged authority claims can influence evaluation even without explicit behavioral directives. The authority-only condition achieves a higher observed TSR than the instructions-only condition on both models: 98.9% vs. 92.3% for Qwen2.5-7B, and 56.7% vs. 26.8% for DeepSeek-flash. The DeepSeek results are particularly informative because the instructions-only condition does not already have a near-saturated TSR.
Adding explicit behavioral instructions to the authority-only payload yields little numerical improvement on Qwen2.5-7B (98.9% to 99.3%) and no numerical improvement on DeepSeek-flash (56.7% to 55.1%). Thus, the observed effect is not explained solely by the additional behavioral instructions, delivery channel, or payload length.
Accordingly, we will revise the Abstract to avoid presenting AEF as categorically distinct from instruction-oriented prompt injection. Instead, we emphasize that forged authorization or endorsement can influence evaluation even without explicit behavioral directives. We will include the complete payloads, character-length-matching procedure, sample-selection details, and statistical methods in the revised appendix.