acceptodds
Under review as a conference paper at ICLR 2027

ATTRACK: Attribution-based Taint Tracking with Provable Bounds for Agent Security

Abstract

Language-model agents act on web pages, emails, and files that may contain indirect prompt injections: instructions planted to redirect agents toward an attacker's goals. We propose ATTRACK, a defence that measures how strongly untrusted content influences each proposed tool call. Its influence test deletes untrusted content and scores the same call again, blocking calls whose log-probability drops beyond a threshold. Citation checks, which detect instructions attributed to untrusted sources, stop most attacks early. The influence test is a certified backstop even when the citation checks are deceived: it bounds the probability that the agent takes a malicious action and the defence misses it, relative to its probability under trusted context alone. Across four AgentDojo suites and two models, ATTRACK blocks every observed attack on 194 matched cases per model. Without attacks, it keeps 89% of the undefended agent's task utility on Qwen3.8-27B and 79% on Gemma-4-26B, compared with 51% and 54% for provenance-taint isolation. On Gemma, by contrast, three detectors (spotlighting, MELON, and a prompt-injection classifier) allow attacks in up to 12.9% of cases. An adaptive attacker finds no successful attack against ATTRACK in 832 payloads over 16 Gemma cases, but finds new ones against all three detectors. For half of AgentDojo's attack cases on Qwen and most on Gemma, the bound keeps the chance that the attacker's exact goal call gets past the defence below one in ten thousand.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
prompt injection, LLM agents, agent security, certified robustness, tool use, adaptive attacks, taint tracking

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_PLY4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces ATTRACK which is a runtime defense against indirect prompt injection in tool using agents. Basically, it scores each proposed tool call with and without the untrusted content and blocks the call if its log probability drops by more than a threshold when the untrusted content is removed. It also adds citation checks which ask a helper model to quote the instruction behind each call and block it if the quote comes from untrusted text. Authors evaluated this on 194 AgentDojo pairs per model, ATTRACK reports zero successful attacks with less utility loss. Also, they adds adaptive attacks and an analysis of how often the bound actually helps.

Critical strengths
  1. The bound does not depend on the citation model getting things right so the formal guarantee and the empirical detection are nicely separated. I think this is very useful because even if an attacker fools the citation model the bound can still holds.
Critical weaknesses
  1. I think the main issue is that the certificates don’t say much about whether a full attack goal is safe. Mostly, the strongest bounds are for one exact serialized call. But for instance an attacker can reach the same goal in many ways like for instance different argument values, different formatting of the same call etc. .The authors somewhat mention this but don’t measure how many goals are actually covered. Also, the main runs use greedy decoding but the bound assumes the call is sampled from the scored distribution.
  2. Moreover, the experiments don’t clearly show what each part adds or how the method compares to the closest work. The main grid uses only 2 injection tasks per suite and most importantly without a defense the attack works on qwen in just 4 of 194 pairs so qwen basically only tells us about utility. and the end-to-end adaptive test covers only 16 Banking cases and it uses a different citation model and prompt than the main runs and lastly the provenance taint baseline leaves out key parts of CAMEL (per-tool policies, user approval, label propagation).
Questions for authors
  1. Can the certificates be extended to cover a full attacker goal (including other calls and equivalent formats of the same call)? What fraction of the tested goals gets a useful bound?
  2. How do you check that the action tokens you score match the ones the model actually generated?
  3. Can you show end-to-end results for the influence test alone, the citation checks alone and the full method along with comparisons to CausalArmor and AttriGuard?
  4. Does the adaptive robustness still hold with the same citation model and prompt as the main runs and on suites other than Banking?
1 comment

Thank you for the review. In short: the bound applies to any set of calls, and we now measure it for whole goals (W1, Q1); Table 1 has each layer on its own, CausalArmor, AttriGuard and the official CaMeL on three models, with an end-to-end adaptive attack on all four suites (W2, Q3, Q4); every check now compares the scored call with the generated tokens (Q2).

W1, Q1 (certificates for a whole goal). Proposition 1 bounds the probability of any set of calls, so it applies to a whole goal: every call, with any argument values and formatting, that makes AgentDojo's success check pass (Section 5, Appendix H). Whether the bound is informative depends on how probable the goal is under the trusted context, and we now measure this. Seven of the eight injection tasks succeed with any call that reaches the attacker's destination. For these goals the median log-probability under the trusted context is −49 on Qwen, −79 on Gemma and −44 on Mistral. At τ = 100 the goal-level bound is therefore above one for almost every pair. At this threshold such goals are protected by the citation checks and not by the bound, and Section 6.2 now says so.

The bound does cover these goals at lower thresholds. At τ = 40 it is below 1e-4 for 49% of Qwen pairs, 87% of Gemma pairs and 14% of Mistral pairs. We ran τ = 40 on tools that change state as a new row of Table 1. It costs 6 to 23 points of utility without attack. No static attack succeeds, and the adaptive attack succeeds in none of Mistral's 104 pairs. These figures assume the destination is the call's first argument. Appendix B gives estimates for other positions, which are lower on Qwen (12% at τ = 40).

Greedy decoding. We reran all pairs at temperature 1 with no truncation (Qwen and Gemma). The attack succeeds in 4 Qwen pairs and 39 Gemma pairs undefended, and in none with ATTRACK.

W2 (what each part adds; closest work; evaluation size). Please see Q3 for the ablations and the closest work, and Q4 for the adaptive attack.

  • Qwen is barely attacked. We agree, and we added Mistral Small 3. The static attack succeeds against it in 83 of 194 pairs undefended and in none with ATTRACK. On all 949 AgentDojo pairs it succeeds in 371 undefended and 5 with ATTRACK.
  • Taint vs. CaMeL. We now run the official CaMeL with its AgentDojo policies. No attack we ran succeeds against it. It completes 41%, 46% and 29% of tasks under attack, and ATTRACK completes 71%, 58% and 27%.

Q2 (do the scored tokens match the generated ones?). Every check now logs the token ids and log-probabilities that the agent generated, and aligns the scored call against them (Appendix D). In the runs behind the ATTRACK rows of Table 1, the scored tokens are the generated ones in 99.6% of Qwen's calls and 99.7% of Gemma's. On Qwen the two log-probabilities agree within 0.73 nats. On Gemma the median difference is 0.03 nats and 99% of calls are within 3.3. Using the generated values would change at most 7 of Gemma's 3,562 verdicts, none of them on a call to an attacker's destination. On Mistral the two agree within 0.72 nats on every call.

The check found one gap in our first runs. A later call in a turn with several parallel calls was scored without the calls generated before it. We fixed this for the ATTRACK rows of Table 1. The adaptive attacks and the ρ-only rows on Qwen and Gemma used the earlier scoring, which lowers ρ on such calls by a median of 9 nats on Qwen, so those rows tend to overstate what gets through.

Q3 (end-to-end ablations; CausalArmor and AttriGuard). Table 1 now has all four, run end to end on the 194 pairs and under the adaptive attack.

  • ρ only. The static attack succeeds in 1, 15 and 27 pairs on Qwen, Gemma and Mistral, and the adaptive attack in 2, 6 and 11 of 40 pairs.
  • Citation only. No static attack succeeds. The adaptive attack succeeds in 0 of 40, 0 of 40 and 2 of 104 pairs. It has no bound behind it.
  • CausalArmor and AttriGuard keep more utility than ATTRACK. On Mistral the static attack succeeds against them in 11 and 6 pairs, and the adaptive attack in 23 and 19 of 104 pairs, compared with 2 of 104 for ATTRACK (Fisher exact p < 1e-3 for each). CausalArmor follows the public causal-armor package; AttriGuard is ported from the authors' artifact.

Q4 (adaptive robustness with the main citer and prompt, beyond banking). The adaptive attack now uses the main citer and prompt on all four suites, end to end, on three models. It succeeds against ATTRACK in none of 40 Qwen pairs, none of 39 Gemma pairs and 2 of 104 Mistral pairs, one of them by replaying a payload found against citation only. Both are bank transfers with ρ below τ (Section 6.2, Appendix A). Two further adaptive attacks on Mistral's 104 pairs succeed in none: one whose objective targets the citation checks directly, and one against τ = 40 on tools that change state.

Reviewer_Vr3XOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ATTRACK, a defense against prompt injection that looks at how much untrusted context raises the chance of a proposed tool call, along with checks on the instruction source. It provides a conditional bound on malicious calls being generated and accepted, relative to trusted context only. In AgentDojo tests on two models, it stopped all observed attacks, including adaptive searches, but at the cost of lower task completion on normal tasks.

Critical strengths
  • The defense is motivated pretty clearly: it measures how strongly untrusted content affects a proposed action, rather than depending only on the model’s guess about whether the content is malicious.

  • The probability-ratio test offers a precise conditional guarantee that still holds even if the auxiliary citation checks get fooled. The paper also points out when that bound becomes uninformative.

  • The evaluation covers both security and task performance, includes adaptive attacks, and looks at how the individual checks work together. That makes it easier to understand both the benefits and the practical costs.

Critical weaknesses

Here are my main concerns:

  • The practical value of the influence test is still pretty unclear. For example, legitimate tasks can rely a lot on tool outputs, which can lead to false positives. And if the injected action looks a lot like the user’s task, let's say, both involve sending an email, the malicious call may already seem plausible in trusted context, so the influence test adds less signal and the citation checks end up doing more of the work. Slo it's not clear that this is a promising defense in production env.

  • The evaluation feels too limited to really show broad robustness. The main AgentDojo experiment only uses two injection tasks per suite, crossed with all user tasks, and tests just two models. These pairs don’t necessarily give independent attack diversity either, since the same payload could carry over to multiple user tasks with the same injection goal. Broader attack coverage and more models would make the results a lot more convincing.

  • The utility–robustness tradeoff doesn’t consistently look better than these baselines. On Qwen, spotlighting blocks all four attacks that work against the undefended agent, while still getting 85.6% task utility under attack vs. ATTRACK’s 71.6%. On Gemma, ATTRACK is more robust than MELON, but utility drops from 72.2% to 64.4% to wipe out the last three successful attacks. The strongest utility comparison is against a simplified provenance-taint baseline, not the full CaMeL system.

  • The theoretical guarantee doesn’t add much practical insight. The bound comes straight from the probability-ratio filtering rule and is only really useful when malicious actions are extremely unlikely in trusted context. The reported numbers mostly cover individual serialized calls, not all calls that could help an attacker succeed. So it’s unclear how much the guarantee says beyond the filtering rule itself or how well it explains the defense performance.

Questions for authors

The adaptive evaluation feels too narrow to support a strong robustness claim, and it should be part of the main results. The TAP search only spans 16 banking cases, and the Qwen probe measures fixed target calls rather than full end-to-end attack success. A better evaluation would add more attack strategies, wider task coverage, and larger search budgets. Adaptive attack results should also be reported alongside the main security–utility comparison, using the same defense configuration, instead of as a separate eval with a different citation model. Could you maybe try stronger attacks used in [1]?

Also, there’s too much space around Figure 1.

[1] The Attacker Moves Second: Stronger Adaptive Attacks Bypass Defenses Against Llm Jailbreaks and Prompt Injections

1 comment

Thank you for the review. In short: ATTRACK keeps 75% to 89% of the undefended agent's utility (W1); we add a much more vulnerable model, all 949 AgentDojo pairs and a second attack template (W2); Table 1 now includes CaMeL, CausalArmor and AttriGuard, and the adaptive attack on all four suites (W3, Q1).

W1 (practical value). False positives when tasks depend on tool outputs. Without attack, ATTRACK keeps 89%, 75% and 85% of the undefended agent's utility on Qwen, Gemma and Mistral. CaMeL keeps 43%, 52% and 51%, and taint isolation 51%, 49% and 47%. Most of ATTRACK's cost comes from travel lookups whose argument is a list returned by the previous tool. The agent repeats the blocked call until its step budget runs out. The refined row exempts read-only tools from the influence test and refuses exact repeats. This raises retention to 90%, 82% and 96%, with no static attack succeeding. We designed the two rules on the Table 1 traces; Section 6.3 reports them on the 755 Mistral pairs not used for the design. On real web pages we measured no cost: on the control pages of BrowseSafe++, where every task is about a page the agent fetches, Gemma completes 78% of tasks undefended and 79% with ATTRACK (21% with the PI classifier; Appendix E).

When the injected action resembles the user's. We agree that the influence test does not help in this case, and the citation checks do the work. With the influence test alone, the static attack succeeds in 15 Gemma pairs and 27 Mistral pairs. With the full method it succeeds in none. The two adaptive successes against ATTRACK are of this kind: bank transfers during banking tasks, where the citer was also deceived.

W2 (breadth).

  • A third model. On Mistral Small 3 the static attack succeeds in 83 of 194 pairs undefended, and in none with ATTRACK.
  • All 949 AgentDojo pairs. On Mistral the attack succeeds in 371 pairs undefended, 78 against CausalArmor, 43 against AttriGuard and 5 against ATTRACK. On Gemma it succeeds in 114 of 921 pairs undefended and in none with ATTRACK. In four of the five Mistral successes against ATTRACK, the injection's goal is a sentence in the final answer, which no tool-call defence can block.
  • A second attack template ("tool knowledge"). It succeeds in 96 of 194 pairs on undefended Mistral and in none with ATTRACK. We ran this template only against these two.
  • Temperature 1. The attack succeeds in 4 Qwen pairs and 39 Gemma pairs undefended, and in none with ATTRACK.

We agree that pairs sharing a payload are not independent, and the paper says so next to its significance tests.

W3 (utility–robustness trade-off).

  • Spotlighting. The static attack succeeds against it in 27 Gemma pairs and 105 Mistral pairs, and the adaptive attack in 3, 11 and 31 of 40 pairs.
  • MELON. The static attack succeeds in 2, 2 and 5 pairs, and the adaptive attack in 2 Qwen pairs and 1 Gemma pair.
  • AttriGuard and CausalArmor keep more utility than ATTRACK. On Mistral the static attack succeeds against them in 6 and 11 pairs, and the adaptive attack in 19 and 23 of 104 pairs, compared with 2 of 104 for ATTRACK (Fisher exact p < 1e-3 for each).
  • CaMeL lets no attack succeed and completes 41%, 46% and 29% of tasks under attack. ATTRACK completes 71%, 58% and 27%.

ATTRACK has higher utility than taint on every model, and than CaMeL on Qwen and Gemma and on Mistral without attack.

W4 (what the theory adds). The proposition says that the calls ATTRACK permits carry a guarantee even when every learned component is deceived, the citer included. No detector we compare against has such a guarantee. It also says where to rely on detection instead, and the data agree: the calls the bound leaves uncovered are the ones the influence test alone misses, and both adaptive successes are on such calls.

We agree that the submission's numbers covered single calls only. The revision adds the bound for a whole goal. At τ = 100 the bound does not cover goals fixed only by their destination. It would at a threshold near 40, which costs a further 6 to 23 points of utility without attack (general response 2/2). The adaptive runs also contain 29 Mistral episodes where the influence test stopped a call that the deceived citer had passed.

Q1 (adaptive evaluation in the main results; stronger attacks). The adaptive attack is now a column of Table 1. It is our implementation of the search-based attacker of Nasr et al., whose code is not released. To fit our compute, we use one search family, 50 payloads per pair, and the agent model as the mutator. It runs on all four suites, end to end, on a sample of 40 pairs per model (104 on Mistral for four defences). On Mistral it succeeds in 28 of 40 pairs undefended and in 31 of 40 against spotlighting. Against ATTRACK it succeeds in none of 40 Qwen pairs, none of 39 Gemma pairs and 2 of 104 Mistral pairs.

Q2 (Figure 1 spacing). Fixed.

Reviewer_7frfOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ATTRACK, a defense against IPI in tool-using LLM agents that measures the causal influence of untrusted context spans on proposed tool calls. ATTRACK couples two mechanisms: (1) an auxiliary LLM instruction-citation check to intercept low-influence preparatory attack steps early, and (2) an action-influence test (ρ\rho) that teacher-forces and scores proposed tool calls after ablating all untrusted tool outputs and previous agent messages, blocking any call where the log-probability drops by more than τ\tau. Supported by a mechanically verified Lean 4 proof, the authors show that bounding ρ≤τ\rho \le \tau bounds the probability of executing an injected action relative to its baseline under trusted context alone. Across AgentDojo suites and Tree-of-Attacks-with-Pruning (TAP) adaptive searches on Qwen3.8-27B and Gemma-4-26B, ATTRACK achieves 0% Attack Success Rate while preserving 71.6%–79.4% task utility.

Critical strengths
  • Bounding the log-probability ratio (ρ≤τ\rho \le \tau) under the agent's own weights directly couples an attack's success probability to its detectability without assuming an infallible sanitizer or external judge.
  • Effective Hybrid Division of Labor: Pairing the certified action-influence backstop (ρ\rho) with empirical instruction-citation checks resolves the core dilemma of attribution defenses: τ\tau can be set high (τ=100\tau = 100 nats) to accommodate legitimate data-dependent arguments, while citation checks catch low-ρ\rho preparatory steps.
  • Achieves 0% ASR across static benchmarks and adaptive white-box searches while retaining significantly higher utility (71.6%–79.4%) than naive taint tracking (43.3%–48.5%), backed by mechanically verified Lean 4 proofs.
Critical weaknesses
  • Proposition 1 assumes temperature-1 ancestral sampling over the entire equivalence class MM of malicious calls. In contrast, the evaluation uses deterministic greedy decoding and evaluates bounds on a single ground-truth serialized string, which can overestimate protection due to surface-form token entropy.
  • At τ=100\tau = 100 nats, the theoretical bound is uninformative (eτp(a∣c∖U)≥1e^\tau p(a \mid c \setminus U) \ge 1) for 20%–51% of attack pairs and for short, low-entropy actions (e.g., password resets), where security relies entirely on the uncertified auxiliary citer.
  • Ablating all prior assistant messages scores a step-tt action as if it were step 1. Conflating trajectory-ordering surprise with untrusted influence inflates ρ\rho for benign multi-step workflows, forcing the threshold to remain high (τ=100\tau = 100 nats).
  • The adaptive evaluations alter the citation prompt, switch the auxiliary citer (using Gemma to cite itself rather than Qwen), and rely on single-step forced-continuation scoring on Qwen rather than end-to-end episode rollouts.
  • Requires arbitrary forced-continuation prompt_logprobs (incompatible with standard closed APIs, while proxy-model scoring shows near-zero rank correlation), and sequential checks add 1.25s–4.4s latency per tool call.
Questions for authors
  • Lack of Online Component Ablations. What is the empirical ASR and utility of ATTRACK in a live, end-to-end ablation using only the certified action-influence test (ρ≤τ\rho \le \tau) without auxiliary citation checks?
  • How can the formal bound be made non-vacuous for short or contextually plausible malicious tool calls (e.g., length-calibrated thresholds or separating function-name influence from argument influence)?
  • Does Gemma's 0% adaptive evasion rate hold under full episode rollouts using the main-table setup (Qwen as citer with the original prompt)?
1 comment

Thank you for the careful review. In short: at temperature 1 no attack succeeds against ATTRACK (W1); the bound is now measured for whole goals, and we agree it does not cover short or plausible calls at τ = 100 (W1, W2, Q2); the adaptive attack and the ablations now run end to end in the main configuration (W4, Q1, Q3).

W1 (greedy decoding; one serialized string). Every call ATTRACK permits satisfies p(a|c) ≤ e^τ p(a|c\U) under any decoder. Proposition 1 turns this into a probability bound when calls are sampled from the scored distribution. We reran all pairs on Qwen and Gemma at temperature 1 without truncation. The attack succeeds in 4 undefended Qwen pairs and 39 undefended Gemma pairs, and in none with ATTRACK.

On the equivalence class: Proposition 1 holds for any set M of calls, including every serialization and argument choice that reaches the goal (Section 5). The revision also measures the bound for a whole goal, any call that reaches the attacker's destination. Such a goal is much more probable under the trusted context than one exact call (median log-probability −49 on Qwen, −79 on Gemma, −44 on Mistral). At τ = 100 the bound therefore does not cover it, and the paper says so (general response 2/2; Section 6.2, Appendix B).

W2 (τ = 100 leaves short or plausible actions uncovered). We agree. No influence test can certify a call that the user's request already makes likely. The revision adds three things:

  1. The ablation shows that these are the attacks the influence test alone misses: 1, 15 and 27 pairs on Qwen, Gemma and Mistral, respectively. The citation checks stop all of them under the static attack.
  2. The adaptive attack succeeds against ATTRACK in 2 of the 183 pairs attacked, both Mistral transfers with ρ below τ (61 and 74) that also passed the citation checks.
  3. A lower threshold on tools that change state covers many of these goals. At τ = 40 the bound on a destination-only goal is below 1e-4 for 49% of Qwen pairs and 87% of Gemma pairs when the destination is the call's first argument (12% and 84% by our estimate for any position). This setting blocks both Mistral payloads and costs a further 6 to 23 points of utility without attack (the τ = 40 row of Table 1; Appendix B).

W3 (deleting earlier agent turns). We measured both sides on Gemma's undefended attack traces. If the agent's own turns were trusted, 7 of the 27 attack steps above τ would fall below it, because the agent had already turned toward the injected call in an earlier turn of its own. For the other calls of those episodes, deletion raises ρ by a median of 21 nats.

Most of ATTRACK's utility cost has a different source: travel lookups whose argument is a list of names returned by the previous tool. The refined row exempts read-only tools from the influence test (the citation checks still apply) and refuses exact repeats of a blocked call. The bound on every call that changes state is unchanged. No static attack succeeds, and utility without attack rises from 80.4%, 67.4% and 46.4% to 81.4%, 73.7% and 52.6% (Section 6.3). We have not run the adaptive attack against this variant.

W4 (adaptive evaluation in a different configuration). The adaptive attack now uses the main configuration: Qwen as citer for every model, the main prompt, and end-to-end episodes on all four suites. One scoring difference remains for later calls of a turn with several calls (Appendix D).

W5 (log-probability access and latency). The submission reported both in the appendices. Section 7 now lists them as limitations. ATTRACK needs the agent's own log-probabilities, and a proxy model's scores do not substitute for them (Appendix I). A check takes a median of 1.25 s on Qwen and 2.33 s on Gemma (Appendix G).

Q1 (ablations). Please see the rows "ρ only" and "citation only" in Table 1, and "What each layer adds" in general response 2/2. With ρ only, the static attack succeeds in 1, 15 and 27 pairs, and the adaptive attack in 2, 6 and 11 of 40 pairs. With citation only, no static attack succeeds, and the adaptive attack succeeds in 2 of 104 Mistral pairs.

Q2 (making the bound non-vacuous for short actions). Per-tool thresholds help when benign calls to a tool are themselves predictable: they raise the share of Qwen's reference calls bounded below 1e-4 from 49% to 63%, and make no difference on Gemma (Appendix B). Separating the function name from its arguments does not help in the hard case. A password reset is plausible in name and short in arguments, so neither part is improbable under the trusted context. For such calls ATTRACK relies on the citation checks, and the two Mistral transfers show that these checks can fail.

Q3 (Gemma adaptive with the main setup). Yes. With Qwen as citer, the main prompt and full episodes, the adaptive attack succeeds in none of the 39 Gemma pairs. Undefended, it succeeds in 18 of 35.

Comments on the paper

2 comments

We thank the reviewers for their comments. The main changes in the revised PDF are:

  1. One comparison table (Table 1 of the revised paper, reproduced below). 13 defences and variants on three models: static attack success, utility with and without attack, and an adaptive attack run end to end in the main configuration.
  2. New baselines and a new model. We add the official CaMeL, CausalArmor, AttriGuard, and each of ATTRACK's two layers on its own. We add Mistral Small 3 (Mistral-Small-24B-Instruct-2501), against which the static attack succeeds in 83 of 194 pairs when undefended.
  3. Main result. The static attack never succeeds against ATTRACK. The adaptive attack succeeds against it in 2 of the 183 pairs attacked, both on Mistral. On the same 104 Mistral pairs it succeeds against AttriGuard in 19 and CausalArmor in 23.
  4. What the bound covers. We now measure the bound for a whole attacker goal, not only for one exact call (comment 2/2).

Setup. The main comparison uses the same 194 injection pairs per model as the submission: every user task of the four AgentDojo suites, each with two injection tasks. Gemma's chat template cannot render two workspace user tasks for scoring, so every Gemma row uses the remaining 190 pairs. All runs were redone on one serving deployment, so some numbers differ slightly from the submission's.

The adaptive attack follows the search-based attacker of Nasr et al. (2025). It rewrites the injection, runs each candidate as a full defended episode, and counts a success only when the task's own check passes. Each defence is attacked on a fixed sample of 40 pairs (10 per suite) with 50 payloads per pair. On Mistral we extended the sample to 104 pairs for four defences; the extra 64 pairs were fixed before these runs.

Table 1a: successful attacks (lower is better). Static successes out of 194 pairs (190 on Gemma) / adaptive successes out of the pairs attacked.

QwenGemmaMistral
undefended3 / 3 of 2833 / 18 of 3583 / 28 of 40
spotlighting0 / 3 of 4027 / 11 of 40105 / 31 of 40
PI classifier0 / 0 of 403 / 3 of 3210 / –
MELON2 / 2 of 402 / 1 of 405 / –
taint isolation0 / 0 of 400 / 2 of 400 / –
CaMeL0 / 0 of 280 / 0 of 400 / –
CausalArmor0 / 1 of 392 / 3 of 4011 / 23 of 104
AttriGuard0 / 1 of 400 / 3 of 406 / 19 of 104
ρ only (ours)1 / 2 of 4015 / 6 of 4027 / 11 of 40
citation only (ours)0 / 0 of 400 / 0 of 400 / 2 of 104
ATTRACK (ours)0 / 0 of 400 / 0 of 390 / 1 of 104 (2 with a replayed payload)
ATTRACK, τ = 40 (ours)0 / –0 / –0 / 0 of 104
ATTRACK, refined (ours)0 / –0 / –0 / –

Table 1b: utility in % (higher is better). Tasks completed under attack / without attack.

QwenGemmaMistral
undefended81.4 / 90.778.9 / 89.527.3 / 54.6
spotlighting85.1 / 91.876.3 / 88.430.4 / 57.7
PI classifier30.9 / 46.423.2 / 41.119.6 / 29.9
MELON77.3 / 82.574.2 / 85.328.4 / 50.5
taint isolation41.8 / 46.441.6 / 44.220.6 / 25.8
CaMeL41.2 / 39.245.8 / 46.328.9 / 28.1
CausalArmor78.4 / 86.676.8 / 85.337.1 / 50.5
AttriGuard85.1 / 89.775.8 / 83.243.3 / 47.4
ρ only (ours)71.1 / 82.561.6 / 65.329.4 / 50.5
citation only (ours)80.4 / 87.673.2 / 86.333.5 / 54.6
ATTRACK (ours)70.6 / 80.458.4 / 67.426.8 / 46.4
ATTRACK, τ = 40 (ours)56.7 / 62.543.1 / 42.223.2 / 40.2
ATTRACK, refined (ours)74.2 / 81.464.7 / 73.732.0 / 52.6

Notes:

  • A dash is an adaptive attack we did not run for lack of compute (50 full defended episodes per pair). A cell with fewer than 40 pairs is an attack that did not finish on every pair, except ATTRACK on Gemma, whose fortieth pair is one of the two tasks dropped above.
  • "τ = 40" uses the threshold 40 for tools that change state (send, write, delete, pay) and 100 for read-only tools. Its Gemma cells cover 188 of the 190 pairs and 90 of the 95 tasks.
  • "Refined" exempts read-only tools from the influence test and refuses exact repeats of a blocked call (Section 6.3).

What the tables show.

  • Every detection and attribution baseline lets static attacks succeed on at least one model, and the adaptive attack succeeds against spotlighting, MELON, CausalArmor and AttriGuard on every model we ran it on.
  • On Mistral's 104 pairs the adaptive attack succeeds in 23 pairs against CausalArmor and 19 against AttriGuard, compared with 2 against ATTRACK (Fisher exact p < 1e-3 for each). CausalArmor and AttriGuard also attribute calls to their context, but neither bounds what a permitted call can be. ATTRACK's influence test does (Proposition 1).
  • The isolation defences (CaMeL and taint) and the PI classifier stop most or all attacks but give up much more utility. Without attack, ATTRACK completes 80%, 67% and 46% of tasks on Qwen, Gemma and Mistral. CaMeL completes 39%, 46% and 28%, taint 46%, 44% and 26%, and the PI classifier 46%, 41% and 30%.

The two adaptive successes against ATTRACK. Both are bank transfers to the attacker's account that Mistral made during a banking task. One was found by the attack on ATTRACK itself. The other was found by the attack on citation only and succeeded when we replayed it against ATTRACK. In both, ρ was below τ = 100 (61 and 74), and the citer attributed the transfer to the user. A transfer during a banking task is plausible without the injection, which is the case Proposition 1 does not cover, so the attack had to deceive the citer as well. With τ = 40 on tools that change state, both payloads are blocked (Section 6.2, Appendix A).

What the bound covers for a whole goal. The submission reported the bound for one exact call per injection. Seven of our eight injection tasks succeed with any call that reaches the attacker's destination (an account, a URL, an address), whatever the other arguments are. We now score such goals directly (Section 6.2, Appendix B):

  • Under the trusted context, the median log-probability of a destination-only goal is −49 on Qwen, −79 on Gemma and −44 on Mistral. At τ = 100 the bound is therefore above one for almost all of these goals. They are protected by the citation checks and not by the bound, and Section 6.2 now says so.
  • At τ = 40 the bound is below 1e-4 for 49% of Qwen pairs, 87% of Gemma pairs and 14% of Mistral pairs (70% of Mistral pairs at τ = 30). These figures assume the destination is the call's first argument. Appendix B gives estimates for other positions, which are lower on Qwen (12%).
  • We ran τ = 40 on tools that change state (the τ = 40 row). It costs a further 6 to 23 points of utility without attack. No static attack succeeds, and the adaptive attack succeeds in none of Mistral's 104 pairs. Compared with 2 of 104 at τ = 100 this difference is not statistically significant.

We keep τ = 100 as the operating point.

What each layer adds (rows "ρ only" and "citation only").

  • The influence test alone lets short calls through when the trusted context already makes them plausible (invitations, calendar events, transfers). The static attack succeeds against it in 1, 15 and 27 pairs.
  • The citation checks alone stop every static attack, and the adaptive attack succeeds against them in 0 of 40, 0 of 40 and 2 of 104 pairs. The adaptive attack does not separate citation only from the full method statistically, and we do not claim it does.
  • The influence test adds the protection that does not depend on a learned judge. If the citer is fully deceived, citation checks stop nothing, while the influence test still stops 3 of 3, 21 of 33 and 58 of 83 static attacks on Qwen, Gemma and Mistral. In the adaptive runs on Mistral we also found 29 episodes, in 11 pairs, where the citer passed a call to the attacker's destination and only the influence test blocked it. Replayed without the influence test, the call ran in 24 of them (Appendix B).

Breadth checks. Running all 13 rows on the full benchmark was beyond our compute. We ran all 949 AgentDojo pairs for a subset of defences, a second attack template and temperature-1 sampling (reply to Vr3X, W2). We also tested ATTRACK on BrowseSafe++ pages. No page made the attack succeed even against the undefended agent, so that benchmark measures cost only: Gemma completes 78% of the control-page tasks undefended and 79% with ATTRACK.

We did not change τ = 100, the deletion of earlier agent turns, or the statement that the bound says nothing about goals that the user's own task already makes likely.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.