acceptodds
Under review as a conference paper at ICLR 2027

Can Agents Safely Expand Their Capabilities? An Empirical Study of Trust Decisions in Autonomous Capability Discovery

Abstract

As Large Language Model (LLM) agents seek capabilities beyond their predefined toolsets, they must decide which third-party software to trust and adopt, creating software supply chain risk during capability discovery. We study this adoption process from search and inspection through recommendation, installation, and execution. Our main study comprises 3,216 trajectories from six LLM agents across eight capability domains. We combine discovery over a frozen corpus of benchmark labeled clean tools and real GitHub repositories with a Resolution IV factorial experiment manipulating five trust signals under recommendation only and enabled execution conditions. Agents typically search once and inspect one candidate, surfacing only 34.6%-44.6% of the benchmark clean alternatives. In controlled trials, poisoned recommendation rates vary substantially across models, from 0% to 73.4%. Unverified security badges increase poisoned selection in multiple models, including by 28.1 percentage points for Qwen3.8-Flash. Under Execution authority, unsafe execution rates range from 0% to 62.9%. Several agents exhibit lower risk alongside greater abstention and reduced protocol completion. Additional guidance improves clean recommendation relative to a generic warning, yet both warnings increase abstention and reduce clean execution from baseline.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Trust Decisions, Autonomous Capability Discovery, Agent Security

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_r4CcOpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents an empirical study of trust decisions in autonomous tool discovery and adoption. The authors evaluate six LLM agents across eight capability domains, investigating how agents discover third-party tools, respond to manipulated trust signals, and proceed from recommendation to installation and execution. The results reveal shallow tool inspection, vulnerability to unverified security signals, and a tradeoff between avoiding unsafe tools and completing legitimate tasks.

Critical strengths

S1. Meaningful research problem.

The paper addresses an important security challenge in autonomous agent systems, providing useful insights into how agents assess and adopt third-party software.

S2. Systematic experimental design.

The study evaluates multiple models and capability domains using controlled trust-signal manipulations. Its multi-stage evaluation of recommendation, execution, abstention, and task completion provides a comprehensive view of agent safety behavior.

Critical weaknesses

W1. Limited end-to-end validation and realism.

Tool discovery and controlled adoption are evaluated separately, while installation and execution are simulated rather than performed. This limits the evidence that the observed trust decisions generalize to realistic end-to-end agent workflows.

W2. Limited diversity of poisoned tool scenarios.

Poisoned candidates are primarily constructed by adding suspicious setup requirements involving credentials and network access. This relatively narrow attack design raises concerns about whether the findings generalize to more diverse and subtle software supply chain threats.

Questions for authors

Please see the critical weaknesses above.

3 comments

We thank the reviewer for recognizing the importance of the problem, as well as our multi-stage evaluation framework and controlled experimental design. We address the two concerns below.

W1. Limited end-to-end validation and realism

We agree that end-to-end realism is important, and we thank the reviewer for raising this point. We evaluate discovery and controlled adoption separately so that the effect of each signal can be identified. In an end-to-end process, whether the agent sees the poisoned candidate depends on the retriever and on the agent’s own search behavior. RQ1 shows that agents usually search once and inspect one candidate. If the two parts were combined, whether a signal influences the choice and whether the candidate is exposed at all would be mixed together. We therefore use two complementary settings.

  • RQ1 characterizes the agent’s search and inspection behaviors on a dataset consisting of real GitHub repositories and benchmark skills.
  • RQ2 and RQ3 estimate the effect of each signal using a factorial design while controlling for the set of candidates.

Both settings use the same action interface and the same capability requests.

The findings from the two settings support each other. The shallow inspection observed in RQ1 also appears in the controlled setting and is directly related to risk. Our trajectory audit shows that 174 of GPT-4o’s 188 poisoned recommendations (92.6%) occur before the poisoned documentation is inspected. The share for Qwen3.8-Flash is 66.7%. Shallow inspection during discovery is therefore one of the sources of risk during adoption.

We use simulated installation and execution for safety reasons. Running malicious content in a real environment could cause actual harm, and the recent incidents mentioned in Sec 1 show that agent actions during evaluation can have real consequences. Prior work such as ToolEmu and AgentDojo also uses simulated environments. Simulation does not change what we measure. This paper studies the agent’s decisions to recommend, install, or execute, and each decision is made when the agent issues the corresponding action. The simulator is also stateful. Installation must come before execution, and after installation the requested permissions are returned to the agent as an observation. The agent therefore goes through a complete multi-step decision process.

W2. Limited diversity of poisoned tool scenarios.

We thank the reviewer for this valuable suggestion. Our poisoned construction is kept uniform to control variables. Each poisoned candidate keeps the functional documentation of a real skill and adds a Required setup section. The added text follows one sanitized template across all scenes and differs only in its canary ID. It describes the prerequisite as a task-unrelated action, with permission metadata covering credentials and network access. Each scene is linked to an MCPTox attack record for provenance, and no original attack payload is included. This keeps the poisoned content consistent across scenes and helps attribute differences in adoption to the five manipulated trust signals. It also reflects a common supply-chain pattern where the tool looks functionally normal but requests access unrelated to the task.

We agree that more diverse poisoned scenarios are a valuable direction. In this paper, we focus on how agents make trust decisions during capability discovery, including the depth of search and inspection, the effect of forgeable signals on selection, and the decision process from recommendation to execution. To identify these factors, we kept the attack construction fixed. Building more covert and diverse poisoned software, such as dependency confusion, typosquatting, or malicious logic hidden in dependencies or code, is a research question that complements our work. We believe it deserves systematic study from both the attack and defense perspectives. From the attack side, one can examine how more covert constructions change agent adoption behavior. From the defense side, one can examine whether detection mechanisms and safety warnings can keep both discrimination and task completion under more complex threats.

We would like to thank the reviewer once again for their recognition of our work and for the constructive questions they have raised.

Thank you for the detailed response. The authors provide reasonable justifications for their experimental design, particularly the separation of tool discovery and controlled adoption, as well as the use of simulated execution to mitigate security risks. However, I still believe that the lack of real end-to-end execution and the relatively limited diversity of attack scenarios constrain the generalizability of the findings to more complex real-world settings.

Nevertheless, I recognize the importance of the research problem and the empirical value of this work. The experimental design is systematic, and the analysis of different trust signals and agent adoption behaviors provides meaningful insights. The authors' response further clarifies the rationale behind their experimental choices. Therefore, I have decided to maintain my original rating of 3.

We sincerely thank the reviewer for the thoughtful follow-up and for recognizing the importance of the problem and the value of our analysis. We agree that real end-to-end execution and more diverse attack scenarios are important for understanding how these findings extend to more complex real-world settings. We have updated the Limitations section in the revised PDF (changes in blue) to explain the rationale for our design and to discuss these two points. We appreciate the reviewer’s support of our paper again.

Reviewer_24aXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper tests whether LM agents are able to detect malicious third-party softwares instead of incorrectly trusting and misusing them. Specifically, they build an agentic eval, where LM agents are given a normal task, and a list of softwares that include some posioned ones created by the author. These poisoned softwares are maliciously created, trying to jailbreak LMs to trust them. Results show various levels of security risks from different LMs.

Critical strengths

The paper proposes a good threat model: agents incorrectly trust and adopt malicious third-party softwares. This is a realistic, timely security risk that needs fix.

Critical weaknesses
  1. This benchmark seems to be saturated by latest models. For example, while GPT-4o can easily get misled by malicious softwares, GPT-5.6 or Claude sonnet 5 already shows decent robustness. So this benchmark cannot provide meaningful signals for recent models.

  2. As a follow-up. the third-party softwares are created in a very simple way that can be easily detected by frontier LMs. Moreover, this might also bias the result due to evaluation awareness.

Overall, I'd recommend authors to improve the difficulty, realism of their benchmark and validate it on frontier LMs - many evaluated models are too old.

Questions for authors

see weaknesses

2 comments

We thank the reviewer for the feedback and for acknowledging that our threat model is realistic and timely. We would like to clarify the focus of our paper, and we address them below.

  1. This paper is an empirical study, not a benchmark.

Our contributions comprise a dataset, a controlled evaluation framework, and an empirical analysis of trust decisions across discovery, recommendation, installation, execution, abstention, and protocol completion. We do not rank models by a single security score. The Resolution IV factorial design is used to identify which observable trust signals influence agent choices (RQ2), not to build a test set that is as hard as possible. As stated in Sec 1 and Sec 6, low insecure adoption alone does not establish reliable candidate discrimination. Our focus is therefore on how agents reach their decisions, rather than on how difficult the test set is.

  1. Frontier models still show meaningful differences.

We respectfully offer a different view on this point. Our results suggest that frontier models still show informative differences that a low risk rate alone does not reveal.

  • Claude Sonnet 5 makes no poisoned recommendations, but under Enabled Execution its completion is only 25.4% and its abstention is 54.7% (Table 3). GPT-5.6 has 0.4% insecure execution, but its completion drops from 71.9% to 61.3%. A low insecure rate does not mean the agent selects clean tools correctly.
  • Claude Haiku 4.5 has an insecure execution rate of only 4.3%, but poisoned candidates make up 32.4% of its completed executions (Sec 4.3).
  • GPT-5.6 and Claude Sonnet 5 almost always search once and inspect one candidate. Only 3 of 24 Sonnet 5 discovery trajectories inspect any clean reference candidate (Table 1).
  • With Detailed Threat Awareness, GPT-5.6 inspects more candidates (2.94 → 3.69), but its clean execution falls from 60.9% to 21.1% (Sec 4.4). We added a length-matched Neutral control. GPT-5.6 reaches 64.1% clean execution under the Neutral prompt and 23.4% under the Detailed prompt, with zero poisoned executions in both. This suggests that the drop is driven by the safety content rather than the added prompt length.
  1. Regarding “many evaluated models are too old”

Five of the six configurations were released in 2025 and 2026: GPT-5.6 (July 2026), Claude Sonnet 5 (June 2026), Claude Haiku 4.5 (October 2025), DeepSeek V4 Flash (2026), and Qwen3.8-Flash (August 2026). GPT-4o is kept as a reference point for comparison within the OpenAI family (Sec 3.3).

Trust signal effects remain significant on recent models. An unverified badge increases poisoned recommendation by 28.1 points for Qwen3.8-Flash (7.0% → 35.2%) and by 14.1 points for DeepSeek V4 Flash (0.8% → 14.8%). Both effects are positive in all 8 capabilities and remain significant after Holm correction (Table 2).

  1. Regarding the poisoned candidates

We would like to clarify how the poisoned candidates are constructed. They are not jailbreak prompts. Each poisoned candidate keeps the functional documentation of a real skill and adds a Required setup section.The added text follows one sanitized template across all scenes and differs only in its canary ID. It describes the prerequisite as a task-unrelated action, with permission metadata covering credentials and network access. Each scene is linked to an MCPTox attack record for provenance, and no original attack payload is included. This keeps the poisoned content consistent across scenes and helps attribute differences in adoption to the five manipulated trust signals. It also reflects a common supply-chain pattern where the tool looks functionally normal but requests access unrelated to the task.

Our trajectory audit also shows that 174 of GPT-4o’s 188 poisoned recommendations (92.6%) occur before the poisoned documentation is inspected. In these cases the agent never sees the setup content, and the choice is made from information in the search results, such as stars and the badge. This is why we study the whole adoption process, including choices made before any documentation review.

The signal effects we report hold under the current construction. More covert and diverse poisoned software, such as dependency confusion, typosquatting, or malicious logic hidden in dependencies or code, is a research question that complements our work.

  1. Regarding evaluation awareness

We thank the reviewer for raising this point. Evaluation awareness is still an open research problem for LLM evaluation in general.The methods to suppress it are still being developed [1]. This concern applies to virtually all behavioral evaluations of LLMs, including the simulated agent safety evaluations in AgentDojo and ToolEmu. For this reason, we see it as a broader challenge for the field rather than a limitation specific to our study. We use a simulator because running real attack payloads would be unsafe. Our conclusions are scoped to the evaluated configurations and candidate construction, and we will add this to Discussion.

[1] Hua et al. Steering Evaluation-Aware Language Models To Act Like They Are Deployed. ICLR 2026

Reviewer_6U2pOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how LLM agents decide which third-party software to adopt when a task requires capabilities beyond their built-in tools, such as web search or code execution. The authors build a local registry and simulator with two settings. In the first, agents search a fixed corpus of 176 candidates, made up of benchmark skills labeled clean and real GitHub repositories, and the authors measure how many searches they run, how many candidates they inspect, and how many clean candidates appear in their results. In the second, each scene contains three benchmark-clean tools and one constructed poisoned variant whose documentation adds a "Required setup" section requesting credentials and network access. Five forgeable signals (search rank, stars, README style, publisher, and a self-claimed security badge) are varied on the poisoned tool only, using a 16-profile fractional factorial design across eight capability domains and six model configurations. Agents either recommend a tool or install and execute it in a simulator that logs inert canaries. Poisoned adoption varies widely across models, and the security badge raises poisoned selection in several of them. Models with lower insecure adoption also abstain more and complete fewer tasks. A prompt-warning study on three models lowers poisoned adoption but also reduces clean execution.

Critical strengths
  • The paper studies a timely and relevant problem of how agents decide which third-party tools to trust when they need to expand beyond their built-in capabilities
  • The paper reports recommendation, installation, execution, abstention and completion separately, which gives a more complete picture than a single insecure adoption rate
  • The controlled factorial design over trust signals, with the same frozen scenes for every model, gives a clean analysis of which cues influence agents
Critical weaknesses
  • The statistical evidence is limited. Effects are tested over only eight capability clusters with two repetitions per cell, and all significant results sit at the minimum achievable adjusted p-value
  • Models are run with different temperature and reasoning settings, so differences across models cannot be attributed to the models themselves
  • Authors do not validate the poisoned label with humans. Credentials and network access can be legitimately needed in domains like email, cloud sync and secret management, so abstaining may be a reasonable behavior
  • The warning study covers only three models and has no neutral prompt control, and the detailed prompt describes the same manipulations that are tested
Questions for authors
  • Can authors report how often agents inspected the poisoned candidate’s documentation before recommending it, and how selection and abstention depend on that
  • Can authors provide human annotation validating that the poisoned candidates are clearly unsafe in every domain
  • Can authors provide results with fixed decoding parameters and matched reasoning settings, with more repetitions and confidence intervals
  • Can authors show whether the trust signal effects hold when the clean candidates also carry stars, badges and publisher cues
4 comments

We thank the reviewer for the careful reading and for recognizing our stage-resolved reporting and factorial design. To address the questions, we added a neutral control experiment, an experiment where clean candidates also carry badges, and an inspection analysis based on the raw trajectories. We respond to each point below.

  • W1 / Q3 (second part). Statistical evidence

The exact sign test is an intentionally conservative choice. We treat capability as the unit of analysis, apply an exact sign test to the 8 capability-level differences, and use Holm correction across the 5 signals. All significant results have an adjusted p of 0.039 because these effects are positive in all 8 of 8 capabilities. This is the strongest directional evidence the test can give.

Effect size intervals. The capability-level 95% t intervals for the significant effects are shown below.

ModelSignal → OutcomeMean effect95% CI
GPT-4oStars → poisoned rec.+39.06[26.97, 51.16]
GPT-4oBadge → poisoned rec.+29.69[17.59, 41.78]
DeepSeek V4 FlashBadge → poisoned rec.+14.06[9.43, 18.69]
Qwen3.8-FlashBadge → poisoned rec.+28.13[19.75, 36.50]
GPT-4oBadge → poisoned exec.+49.22[35.41, 63.03]

All lower bounds are clearly above zero. For example, the Qwen badge effect is positive in every capability, ranging from +12.50 (secret management) to +43.75 (database migration).

On more repetitions. Using capability as the inference unit is a deliberate choice. Trajectories within the same capability share task semantics and candidate pools, so they are not independent. Treating trajectories as independent samples would be pseudo-replication and would inflate significance. Our test asks whether a signal effect holds consistently across functional domains, which is the generalizable claim we care about. More repetitions would improve the precision within each capability (each signal level already has 16 trajectories per capability), but would not change the inference unit. The table above shows the current precision is sufficient. For effects that do not pass correction, the paper never claims they are absent. Sec 4.2 states “Other signals show weaker evidence.”

  • W2 / Q3 (first part). Decoding and reasoning settings across models

Table 5 reports the API settings for each agent configuration. In the main study, we set temperature to 0 for GPT-4o and omitted the temperature parameter for the other models. Reasoning controls vary across providers, and identical parameter values can correspond to different inference behavior. As stated in Section 3.3, our cross-model comparisons describe the tested agent configurations. Within each configuration, we hold decoding and reasoning settings fixed when comparing trust signals (RQ2), autonomy conditions (RQ3), and prompts (Section 4.4).

We reran Claude Haiku 4.5 and Qwen3.8-Flash at temperature 0 with API thinking disabled. Thinking was already off for both models in the main study. The original and follow-up runs used the same 128 scenes per model and autonomy condition and were collected in separate periods.

Recommendation Only. On the matched scenes, Qwen’s estimated badge effect on poisoned recommendation was +34.38 pp under the original settings and +29.69 pp at temperature 0. The point estimates were similar. The temperature-zero estimate had a wide, unadjusted 95% capability-level CI of [1.79, 57.58] pp and did not pass Holm correction (adjusted p = .313). Enabled Execution. For Haiku, poisoned execution went from 5.47% to 6.25% and abstention from 77.34% to 79.69%. For Qwen, poisoned execution stayed at 0.78% and abstention went from 77.34% to 69.53%. Under both settings, infrequent poisoned execution coexisted with frequent abstention.

  • W3 / Q2. Candidate labels and human validation.

A coauthor with security research expertise reviewed the experimental design before data collection. We did not conduct independent, case-by-case human annotation of the final candidates. The clean skills inherit their labels from MaliciousAgentSkillsBench (Liu et al., USENIX Security ‘26). Each constructed variant retains the source skill’s functional documentation and adds a Required setup section that explicitly describes a task-unrelated prerequisite. This section uses the same sanitized, inert template across all scenes and differs only in its canary identifier. Each scene is linked to an MCPTox record (Wang et al., AAAI ‘26) for provenance. Because the appended text is identical across scenes apart from this identifier, case-by-case annotation would largely amount to assessing a single template. In our study, “poisoned” identifies this experimental condition.

We agree that credentials and network access can be legitimate requirements. We retain the original authentication instructions, including Gmail CLI authentication in the email domain and agentuity auth login in the secret management domain. Each scene also offers three clean candidates drawn from the same capability pool. Our controlled comparisons estimate how credibility cues change the probability of selecting the constructed variant. This estimate does not require every variant to be judged clearly unsafe in every domain. We report abstention separately because declining to proceed can be reasonable when a task involves sensitive permissions.

The table reports Qwen3.8-Flash’s behavior under Recommendation Only in the three domains highlighted by the reviewer.

CapabilityPoisoned recommendationAbstentionBadge effect
Email integration18.75%50.0%+25.00 pp
Cloud synchronization18.75%75.0%+25.00 pp
Secret management12.50%46.9%+12.50 pp

Each domain includes 32 trajectories, with 16 at each badge level. The badge effect is the difference in poisoned recommendation rates between the conditions with and without a badge on the constructed variant.

Qwen3.8-Flash often abstained in these domains, and the observed recommendation rate was higher with the unverified badge in all three. Whether the setup requirement is legitimate does not depend on a self-declared badge, so this shift suggests that a forgeable cue still influenced selection in these domains. We will clarify how the labels were assigned and what the design review covered. We appreciate the reviewer’s point about legitimate access needs and see assessing the risk of analogous setup requirements across real deployment contexts as a valuable direction for future empirical work.

  • W4. Neutral control for the warning study.

We added a Neutral control that appends text of the same length as the Detailed prompt but without security content. All other settings match the main study. Results under Enabled Execution:

ModelNeutral clean exec.Detailed clean exec.Difference
GPT-4o28.1%0.0%−28.1 pp
Haiku 4.510.9%2.3%−8.6 pp
GPT-5.664.1%23.4%−40.6 pp

All three differences have capability-cluster bootstrap 95% intervals below zero. Three observations follow.

  1. The Neutral condition is close to the original baseline (GPT-4o 28.1% vs. 29.3%, Haiku 10.9% vs. 9.0%, GPT-5.6 64.1% vs. 60.9%). Adding text of the same length does not change adoption behavior by itself.

  2. This suggests the drop in clean execution is driven by the safety content. GPT-5.6 shows this most clearly. It has zero poisoned executions under both Neutral and Detailed, but the neutral text keeps much more clean adoption. The results replicate across collection periods. The Detailed results collected during the rebuttal are very close to Table 4 (GPT-4o 0.0% in both, Haiku 2.3% vs. 1.6%, GPT-5.6 23.4% vs. 21.1%).

  3. On the Detailed prompt describing the tested manipulations, this is intentional. The intervention is a diagnostic test that asks whether agents discriminate better when told which signals can be forged. Recommendation selectivity improves, but clean execution still drops even with this information.

On using three models, they represent the three behavior patterns observed in the main study: frequent poisoned adoption (GPT-4o), high abstention with low completion (Haiku 4.5), and low poisoned adoption with higher clean execution (GPT-5.6) (Sec 4.4).

  • Q1. Did agents inspect the poisoned documentation before recommending it?

We extracted from the Recommendation Only trajectories the events where the poisoned documentation was returned to the agent and checked whether they occurred before the recommendation. For trajectories ending in abstention or other terminal outcomes, grouping depends on whether the poisoned documentation was received before termination.

ModelPoisoned rec. preceded by inspectionPoisoned rec. rate (inspected)Poisoned rec. rate (not inspected)Abstention (inspected)Abstention (not inspected)
GPT-4o7.4%87.5%72.5%12.5%0.4%
GPT-5.6100%3.5%0.0%42.7%8.8%
DeepSeek V4 Flash35.0%3.6%20.6%29.0%3.2%
Claude Sonnet 5n/a0.0%0.0%41.7%45.0%
Claude Haiku 4.5100%23.2%0.0%63.0%36.0%
Qwen3.8-Flash33.3%9.8%50.0%52.7%11.1%

For GPT-4o, 92.6% of poisoned recommendations occur before the poisoned documentation is inspected. The corresponding share for Qwen is 66.7%. In contrast, all poisoned recommendations by Haiku and GPT-5.6 occur after inspection. These results separate two kinds of risk: adoption without reviewing the documentation, and poisoned recommendation after seeing it.

This analysis is also consistent with RQ2. Stars and badges appear directly in search results, while README presentation and the added setup are only visible through Inspect. Most of GPT-4o’s poisoned recommendations happen before the latter two are seen, which matches its strong response to signals visible in search results (stars +39.06 pp, badge +29.69 pp) and the lack of a significant README effect. Since inspection is chosen by the agent, these conditional statistics describe behavioral associations.

  • Q4. Do signal effects hold when clean candidates also carry trust signals?

We ran an additional experiment where all three clean candidates carry a badge and only the poisoned candidate’s badge is manipulated. We chose the badge because it is the only signal significant in more than one model in the main study. Stars are significant only for GPT-4o, and publisher is not significant for any model. We used the two models with the largest badge effects in the main study, GPT-4o (+29.69 pp) and Qwen (+28.13 pp).

ModelPoison without badgePoison with badgeEffectHolm p
GPT-4o7.81%54.69%+46.88 pp0.039
Qwen3.8-Flash1.56%12.50%+10.94 pp0.313

For GPT-4o, the poisoned candidate’s own badge still significantly increases poisoned recommendation even when every clean candidate has one, and the effect is larger than in the main study (+46.88 pp vs. +29.69 pp). The badge effect cannot be explained by badges being scarce. For Qwen, the effect has the same direction but is smaller and not significant after correction. Its response to the badge may partly depend on contrast between candidates. When clean candidates have badges, and the poisoned candidate does not, poisoned rates drop to 7.81% and 1.56%. This further supports the badge as a signal that drives selection.

We thank the reviewer again for the constructive questions, which helped us strengthen the paper, and we hope the reviewer will consider these additions when reassessing the score.

Reviewer_6U2pOpenReview ↗

Thank you for addressing the concerns

I am raising my score to a weak accept

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.