arXiv 2605.10834v1May 11, 2026

From Controlled to the Wild: Evaluation of Pentesting Agents for the Real-World

Pedro Conde et al.

Brief context

Publication timing, weekly edition context, and source links for this brief.

Published

May 11, 2026, 4:50 PM

Current score

80

Original paper

The executive brief below is grounded in the source paper and linked back to the arXiv abstract.

AI pentesting agents are increasingly credible as offensive security systems, but current benchmarks still provide limited guidance on which will perform best in real-world targets. Existing evaluation protocols assess and optimize for predefined goals such as capture-the-flag, remote code execution, exploit reproduction, or trajectory similarity, in simplified or narrow settings. These tools are valuable for measuring bounded capabilities, yet they do not adequately capture the complexity, open-ended exploration, and strategic decision-making required in realistic pentesting. In this paper, we present a practical evaluation protocol that shifts assessment from task completion to validated vulnerability discovery, allowing evaluation in sufficiently complex targets spanning multiple attack surfaces and vulnerability classes. The protocol combines structured ground-truth with LLM-based semantic matching to identify vulnerabilities, bipartite resolution to score findings under realistic ambiguity, continuous ground-truth maintenance, repeated and cumulative evaluation of stochastic agents, efficiency metrics, and reduced-suite selection for sustainable experimentation. This protocol extends the state of the art by enabling a more realistic, operationally informative comparison of AI pentesting agents. To enable reproducibility, we also release expert-annotated ground truth and code for the proposed evaluation protocol: https://github.com/jd0965199-oss/ethibench.

Score 80Full-paper briefagentsmodelsinferencedata

Executive brief

A short business-reader brief that explains why the paper matters now and what to watch or do next.

Why this is worth your attention

AI pentesting agents are getting credible enough that the bottleneck is no longer just capability—it is knowing which systems actually find real vulnerabilities without drowning teams in noise, duplicates, cost, and irreproducible results. This paper offers a practical evaluation recipe that looks much closer to how security teams buy and operate tools: validated findings, repeated runs, cost and runtime, severity, coverage, and false-positive control. The evidence is useful but not a final vendor leaderboard; it is a signal that security, procurement, and platform teams should start demanding operational evaluations rather than demo-friendly exploit benchmarks.

  • The paper’s strongest contribution is not a new hacking agent; it is a better buying and testing frame. Ask for validated vulnerability findings, duplicate handling, false-positive rates, and severity coverage—not capture-the-flag wins or impressive-looking attack traces.
  • A single successful run is a weak signal for agentic pentesting because small model-output changes can cascade through long tool chains. Vendors should show repeated-run means and variance, total runtime, monetary cost, discovery over time, and what happens when findings are accumulated across runs.
  • The experiments suggest a real operating trade-off: higher recall can come with duplicates and noise, while cleaner outputs may miss more vulnerabilities. The same engine can also change materially when paired with a different model backend, so vendor comparisons need to lock both the agent and the underlying model.
  • If your organization is serious about AI-assisted security testing, the useful move is to create a small but realistic target suite with expert-maintained ground truth and a cheaper reduced version for frequent experiments. Without that, teams will mostly be comparing demos, not operational value.
  • The empirical base is useful but narrow: three targets, 108 annotated vulnerabilities, three runs per experiment, and a 50-finding triage sample for the matching pipeline. The protocol also does not yet test whether agents avoid destructive actions or how they behave in changing, patched environments.

Evidence ledger

The strongest claims in the brief, along with the confidence and citation depth behind them.

strategichighp.1

The paper proposes a finding-level evaluation protocol for AI pentesting agents rather than relying on task completion, trajectory similarity, or capture-the-flag style success.

stackhighp.3p.3

The scoring pipeline uses LLM semantic matching plus bipartite resolution to match agent findings to ground truth while limiting duplicate credit.

inferencehighp.5p.6

The protocol treats stochasticity, runtime, cost, and cumulative discovery as first-class operational metrics.

caveathighp.9p.9

The work is primarily an evaluation methodology and does not yet cover safety behavior or introduce new benchmark targets.

Related briefs

More plain-English summaries from the archive with nearby topics or operator relevance.

cs.LG

OpenWebRL: Demystifying Online Multi-turn Reinforcement Learning for Visual Web Agents

Rui Yang et al.

cs.CR

Grounded Cache Routing for Retrieval-Augmented Generation: When Is It Safe to Reuse an Answer?

Syed Huma Shah

cs.AI

The Illusion of Multi-Agent Advantage

Prathyusha Jwalapuram et al.

cs.AI

A History-Aware Visually Grounded Critic for Computer Use Agents

Jaewoo Lee et al.

Thank you to arXiv for use of its open access interoperability. This product was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.
LightDark