acceptodds
Under review as a conference paper at ICLR 2027

SuperRed: AI Red-Teaming and Security Benchmarking with a Fine-grained Threat Model

Abstract

Current AI red-teaming evaluations are fragmented across standalone attacks and benchmarks with different threat models, making reported attack success rates (ASRs) difficult to compare and reproduce. For the same attack strategy or benchmark, changing the attacker model and access scope moves ASR from 2.7% to 67.0%. We propose the open-source framework SuperRed for modularized AI red-teaming and security evaluation. SuperRed explicitly specifies and enforces attacker capabilities, including access, feedback, model, and budget, under a common execution model. Attackers, targets, and benchmarks are interchangeable modules, enabling compatible combinations to be evaluated without strategy-specific integration. Using SuperRed, we evaluate compatible combinations of 11 attack strategies, 6 benchmarks, 13 attacker models, and 11 access scopes, using approximately USD 130,000 of inference across chatbot and agent targets. Open-weight models are the strongest tested jailbreak attackers and strong against agent targets (Claude Code). Observing target trajectories provides the largest access advantage; together with judge feedback, it raises AgentDojo ASR from 15.2% to 58.9%. Increasing attacker budget shows sharply diminishing returns after USD 0.19 per jailbreak task and USD 0.25 to USD 0.57 per agent task. These results show that attack success is conditional on the threat model and should be reported and compared accordingly, and that a poorly resourced attacker with an open-weight model already attacks close to the ceiling of current strategies.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
security, agents, jailbreak, prompt injection, threat model, red-teaming

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_i6zoOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces SuperRed, an open-source framework for modularized AI red-teaming and security evaluation. It explicitly specifies and enforces attacker capabilities, including access, feedback, model, and budget, while making attackers, targets, and benchmarks interchangeable modules. Through large-scale evaluations across chatbot and agent targets, the authors show that attack success is highly conditional on the threat model. Open-weight models are the strongest tested jailbreak attackers, observing target trajectories provides the largest access advantage, and increasing attacker budget shows sharply diminishing returns.

Critical strengths
  1. Inconsistent threat models and evaluation protocols are a serious problem in AI red-teaming, as I have also seen in practice. SuperRed effectively addresses this issue through a unified evaluation framework.
  2. The paper involves substantial engineering effort and extensive experiments.
Critical weaknesses
  1. SuperRed achieves fairness mainly through restrictions, which may not cover the needs of new attack methods and could limit flexibility. A simpler framework that clearly documents permissions and evaluation settings might be more practical.
Questions for authors
  1. How does SuperRed support new attack capabilities beyond its current threat model without modifying the core framework?

  2. How is the original behavior of ported attacks preserved, given the reported GEPA integration issues?

  3. How much effort is required to integrate new attacks, and what are the benefits of strict enforcement over simply documenting permissions and evaluation settings?

5 comments
AuthorsOpenReview ↗

Response to W1: flexibility and enforcement vs. documentation

SuperRed achieves fairness mainly through restrictions, which may not cover the needs of new attack methods and could limit flexibility.

SuperRed has flexibility to support the needs of novel attack methods through observables and controllables which satisfy those needs within the framework.

A simpler framework that clearly documents permissions and evaluation settings might be more practical.

  1. Documentation describes one fixed setup. SuperRed's threat model is explicitly parameterized (attacker model, scope, feedback, budget) which allows to evaluate with variable control. As threat model dimensions can be independently changed in a repeatable way, we can understand the impact of each dimension
  2. Scopes use the same kind of tags across targets, so the same parameters apply to every attack and target. This makes results comparable across targets and attacks
AuthorsOpenReview ↗

Response to Q1: new attack capabilities

How does SuperRed support new attack capabilities beyond its current threat model without modifying the core framework?

  1. The threat model is flexible by modeling capabilities as a configurable part. Targets can be configured with new observables and controllables to satisfy new capability requirements within the framework
  2. For example: To support tool-output poisoning, one adds a tool-return surface as a controllable in the target wrapper
  3. No core change is needed if the capability fits the existing event interfaces. We chose events as the most general abstraction because everything from injections to network interactions can be modeled as an event. Very different interaction semantics may need an extension
AuthorsOpenReview ↗

Response to Q2: preserving ported attacks

How is the original behavior of ported attacks preserved, given the reported GEPA integration issues?

  1. If a reference implementation is available for an attack, the module can keep that implementation unchanged
  2. If no reference implementation is available for an attack, an implementation is performed as best-effort to the original concept
  3. If an original attack has a narrow focus, we may extend it to generalize across targets. This adds behavior beyond the original and such additions are documented in the module. GEPA is such a documented adaption: keeps reflective mutation and the meta-prompt; current-best instead of Pareto selection; acceptance test dropped
  4. In general, preservation is checked per port, not guaranteed, with deviations recorded in an ASSUMPTIONS file. Testing, which also revealed the GEPA integration bug, is partially borrowed from upstream code and partially added (such as regression tests against the reflection failures)
Reviewer_wcjFOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

the paper introduces SuperRed, a modular framework for controlled security evaluation. It does large scale evaluation in current red-teaming methods on this framework.

Critical strengths
  • the paper is relevant and the work is suitable for this conference
  • the writing is mostly clear , and the goal of the paper is clearly defined
  • the results are very comprehensive and a good analysis
Critical weaknesses
  • missing error bars in the results
  • GPT-5.4 is also a victim and judges its own replies, and those cells weren't re-judged
  • There's no human validation of the judge
  • Unreported dimensions are averaged with equal weight, across benchmarks whose ASR levels differ greatly
Questions for authors
  • How well does the GPT-5.4 judge agree with each benchmark's original classifier and with human labels? How do the ASRs and rankings change with the original judges, or with a judge that isn't also a victim?
  • How were the 79 chatbot goals, and the 18 added AgentDojo tasks, selected? Do the conclusions hold on the full HarmBench, StrongREJECT and SORRY-Bench sets?
  • Can you give confidence intervals for the main claims (Tables 3–5), and test the ones that rest on small gaps, e.g. open-weight vs proprietary on DTAP, and strategy rankings?
7 comments
AuthorsOpenReview ↗

Response to W1: error bars

missing error bars in the results

Thank you for flagging this. We will add error bars in the revision

AuthorsOpenReview ↗

Response to Q3: confidence intervals

Can you give confidence intervals for the main claims (Tables 3–5), and test the ones that rest on small gaps, e.g. open-weight vs proprietary on DTAP, and strategy rankings?

  1. Appendix H reports both the reproducibility and the statistical certainty of our results. In three identical repeats of twelve chatbot combinations, no setting deviates from its mean by more than 1.8 pp and no single strategy by more than 5.2 pp. Bootstrap resampling of the goals gives 95% confidence intervals of about ±2 to ±9 pp for the strongest-attack entries of Table 4 and ±2 to ±5 pp for its means
  2. In the revision we will report these confidence intervals in the main text, in Tables 3 to 5 and the figures, and mark comparisons whose gap lies within them
AuthorsOpenReview ↗

Response to W4: averaging across benchmarks

Unreported dimensions are averaged with equal weight, across benchmarks whose ASR levels differ greatly

  1. This is by design. Each table and figure reports two or three dimensions and averages the others with equal weight. A single figure with all dimensions would not be readable, so each figure shows the general trend along the dimensions it reports.
  2. Where benchmarks disagree, we report them separately: Table 5 gives ASR per benchmark, and Section 4.4 discusses where strategy rankings change between benchmarks. We will add the relevant splits in the revision; in the main body or appendix depending on space
  3. We always disclose all aggregations transparently: The captions state which dimensions are averaged, and the appendix gives the full averaging rule.
Reviewer_3J4yOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces SuperRed, a modular red-teaming framework that separates attack optimizers, targets, and security benchmarks while enforcing explicit constraints on access, feedback, attacker models, and inference budgets. It empirically evaluates how these factors affect attack success across chatbot and agent targets.

Critical strengths
  1. The framework design is reasonable and addresses a practical need for more reproducible and controlled red-teaming evaluations.

  2. The overall experimental effort is substantial, covering diverse attack strategies, attacker models, benchmarks, and access configurations.

  3. Runtime access filtering and outcome-based agent evaluation are useful mechanisms for reducing accidental information leakage and misleading success judgments.

Critical weaknesses
  1. Limited conceptual novelty and new insights. The main findings—that attacker models, access permissions, and feedback affect attack success—are largely expected. Quantifying these effects is useful, but the paper does not clearly demonstrate what previously unavailable insight follows from its unified abstraction. The contribution therefore appears primarily infrastructural and engineering-oriented. Could the authors identify a non-obvious finding or evaluation capability enabled by SuperRed that would be difficult to obtain using existing frameworks?

  2. Extensibility requires further validation, particularly for learning-based attacks. Supporting RL-based methods such as PISmith (Yin et al.) and RL-Hammer (Wen et al.) raises questions beyond inference-time optimization. How would the framework accommodate policy training, weight updates, and training-query costs, rather than only evaluating a frozen attacker? How should training costs be amortized when comparing RL attackers with test-time search? Likewise, integrating a new target requires defining its observables, controllables, and domain mappings. What adaptation effort is required, and how is the implementation validated against omitted attack surfaces or incorrectly granted permissions?

  3. Cross-target scope alignment remains incomplete. The paper explicitly documents that similarly named scopes grant different capabilities across targets. For example, user prompt includes system-prompt modification on AgentDojo but not on DTAP, where this capability belongs to agent environment. Although the framework can represent these differences, the reported scope labels do not establish matched threat models across targets, limiting direct cross-target interpretation. Could the authors explain why this design was especially chosen.?


Wen, Yuxin, et al. "Rl is a hammer and llms are nails: A simple reinforcement learning recipe for strong prompt injection." arXiv preprint arXiv:2510.04885 (2025).

Yin, Chenlong, et al. "Pismith: Reinforcement learning-based red teaming for prompt injection defenses." COLM 2026

Questions for authors

See Weakness

3 comments
AuthorsOpenReview ↗

Response to W1: novelty and insights

Limited conceptual novelty and new insights. The main findings [...] are largely expected. [...] Could the authors identify a non-obvious finding or evaluation capability enabled by SuperRed that would be difficult to obtain using existing frameworks?

  1. SuperRed turns "expected" assumptions into explicit, measurable artifacts: Does an attack's advantage persist with the same model, permissions, feedback and budget? This is a foundational scientific contribution
  2. Creating the infrastructure is definitely an essential part of our work. We then use that infrastructure to perform a comparative analysis that would not otherwise be possible
  3. Some capabilities that existing frameworks lack or support only in part are:
    1. Declaratively grantable access scopes: a target declares what it exposes as trust domains and a threat model grants any subset for reading or writing
    2. Runtime enforcement across dimensions without per-setting code: the controller filters reads and writes to the granted scope, withholds judge feedback when it is off, fixes the attacker model, and evaluates spend per attack
    3. One attack module across attacks, separating strategy from model performance: e.g., each of the five model-driven chatbot strategies ran under seven attacker models
    4. Existing frameworks cover individual parts of this (PyRIT and DeepTeam let the user choose the attacker model, DoomArena configures which component is attackable) but none limits an attack's reads to a granted scope, evaluates the attack cost, or combines all of them
  4. Some findings which we present that are not possible with existing frameworks but SuperRed enables are:
    1. Observation matters more than judge feedback: response access adds 13.3 percentage points to strongest-attack chatbot ASR, versus 1.6 points from feedback alone. (Enabled by allowing separate control over both dimensions)
    2. Attack effectiveness depends heavily on the attacker model: changing the model shifts ASR by 40.2 percentage points on average, holding the strategy fixed. (Enabled by separating the attacker strategy from the attacker model)
    3. More budget does not eliminate access gaps: response-reading attackers exceed blind attackers' capped performance within the tested budgets. (Enabled by attack cost evaluation)
    4. Most measured attacks are cheap: mean ASR reaches 90% of its capped value at 0.19 USD per chatbot task and 0.25-0.57 USD per agent task attacker inference. (Enabled by attack cost evaluation)
    5. High ASR can overstate an attack's contribution: ASB already yields 93.5% success from direct requests, limiting attack effectiveness to improvement over that baseline. (Enabled by comparable evaluation across attacks)
AuthorsOpenReview ↗

Response to W2: extensibility and validation

Extensibility requires further validation, particularly for learning-based attacks. [...] How would the framework accommodate policy training, weight updates, and training-query costs, rather than only evaluating a frozen attacker? How should training costs be amortized when comparing RL attackers with test-time search?

  1. The paper covers 50+ and can't cover the whole field. The community can contribute additional modules to SuperRed. Integrating various environments and tasks are substantial contribution for RL attackers
  2. Although out of scope in the presented modules, RL can be enabled in SuperRed. We can easily point the model proxy endpoint to RL served models. Litellm, which the proxy is based on, supports a finetuning/RL API through which usage can be tracked (we will state this opportunity explicitly in the revision)

Likewise, integrating a new target requires defining its observables, controllables, and domain mappings. What adaptation effort is required [...]?

This is part of our documentation site that is currently hidden for anonymization purposes but also briefly mentioned in the appendix of our paper. We will try to explain it here as best as we can. The following are the steps required for integration:

  1. Identify observable information and attacker injection points. Effort: understanding of the target system
  2. Either wrap the existing system in a reusable target adapter, or add the identified information and injection points in-line in the target code. Effort: About one line per injection point in addition to minimal scaffolding
  3. Define trust domains of the target system and map each information and injection point to one trust domain. Effort: Declarative framework-level function calls as highlighted in Appendix C.2

[...] and how is the implementation validated against omitted attack surfaces or incorrectly granted permissions?

  1. SuperRed focuses on rigorous results by providing infrastructure that can run comparable evaluations. A target may expose only a subset of a real deployment's attack surface and still give comparable results, so exhaustive coverage is not what the framework aims to guarantee by itself
  2. The modules are not part of SuperRed but supplied to it; implementing and validating a target lies with its module author. Each module documents its surfaces, permissions and assumptions explicitly. Since modules are published open source, anyone can check the reasoning
  3. For the modules we authored, we used the following methods that we also recommend to others: We derive surfaces from the target's own interfaces instead of picking them by hand. E.g., in AgentDojo, every tool becomes exactly one controllable tagged at the data store it reads or writes, and tests assert this for every tool. An end-to-end test checks that an access succeeds only with the required scope. We will state this in the revision more clearly
AuthorsOpenReview ↗

Response to W3: scope labels across targets

Cross-target scope alignment remains incomplete. [...] For example, user prompt includes system-prompt modification on AgentDojo but not on DTAP, where this capability belongs to agent environment. [...] Could the authors explain why this design was especially chosen?

  1. User-prompt and system-prompt writes are separate permissions in SuperRed. ASB, AgentDojo and DTAP each declare the two prompts as separate trust domains, and a scope can grant either one alone
  2. The difference comes from how we grouped permissions in experiments. Our ASB and AgentDojo runs used one write mode for both prompts, which covers ASB’s two prompt attacks: direct injection into the user prompt and its plan-of-thought backdoor in the system prompt. DTAP grants them separately and puts the system prompt under agent environment (Table 2 caption, Appendix F.1, Table 10)
  3. In the revision we will state this difference clearly while highlighting how this data shows that ASR gains still come from read access and judge feedback. In short: For AgentDojo, attacks that write only the system prompt reach 15.2% when blind and 58.9% when they also read the agent's trajectory and the judge's feedback. The write access is the same in both, so the 43.7 pp gain comes from reading. For DTAP, where both scopes write only the user prompt, the same step raises ASR for every victim: Opus 4.8 from 17.6% to 24.2%, Sonnet 4.6 from 21.2% to 26.1% and Haiku 4.5 from 20.3% to 30.3% (Table 4).

Comments on the paper

6 comments

Dear reviewers and AC, thank you for the time you put into reviewing SuperRed! We replied to each point under the respective review. Looking forward to the discussion and happy to clarify anything further

AuthorsOpenReview ↗

Response to Reviewer 3J4y

Thank you for your response. We would like to emphasize that the central contribution of SuperRed is the framework itself, rather than the individual empirical findings. This is precisely why we submitted the paper to the Infrastructure, Software Libraries, Hardware, Systems, etc. track. Your point that these findings can be obtained through standard ablations or a benchmark’s own baselines is exactly what we want to address with SuperRed. We do not want to rely on each benchmark's own baselines and assumptions and want to bring them onto a level playing field so they can be compared fairly.

As an example of what SuperRed aims to address, consider two successful peer-reviewed attacks with similar ASRs, but one costs 30x more to execute or requires significantly deeper access to the target agent. Yet, ASR alone often becomes the headline metric driving recognition, citations, and popularity, without accounting for these differences. With SuperRed, we want to change that narrative by making such differences explicit. What may seem obvious at first glance may not necessarily hold up empirically.

We also appreciate your clarification regarding target costs and will account for it in the revision.

We respectfully hope that you reconsider your score in light of the paper's intended contribution.

Thank you for the clarification. I agree that, for the infrastructure track, the framework itself is the main contribution, and I appreciate that the paper does show cases where rankings change once access or cost is accounted for (e.g., GOAT vs. Crescendo in Table 14; AgentVigil vs. GEPA agentic in Sec. 4.1). However, judged as infrastructure, extensibility and correctness are central, and my W2 questions remain unanswered:

  • how training-based attackers (e.g., RL) would be supported;
  • how their training costs would be tracked and amortized;
  • how module implementations are validated.

My concern about cross-target comparability (W3) also bears directly on the "level playing field" argument. As documented in Appendix F.1 and Table 10, scopes with the same name grant different capabilities across targets in at least three ways:

  1. user prompt also writes the system prompt on ASB and AgentDojo, but not on DTAP.
  2. agent environment writes the system prompt, tool catalog, skills and host files on DTAP, but only the tool catalog and memory on ASB and AgentDojo; on DTAP it also does not read the trajectory.
  3. DTAP's non-admin and admin tools scopes additionally include the email and messaging bundle.

Table 4 nevertheless reports these targets side by side under the same column headers. For example, agent environment yields 8.5% on AgentDojo but 14.5–30.6% on DTAP, which partly reflects DTAP's additional system-prompt write access rather than a difference in target robustness. The previous response clarified the differences and promised to state them in the revision, but the differences themselves remain. The paper's own experiments therefore do not yet provide the matched threat models across targets that the framework is meant to enable. Addressing this would require either aligning the scope definitions (e.g., re-running AgentDojo with a user-prompt-only scope) or clearly separating non-comparable results in the tables.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.