{"id":"f396342c-217c-49df-ad94-71df60c87da2","arxiv_id":"2610.08773","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Co-evolving tasks, attacks, and agent inside a frozen web world model makes a 4B web agent more capable and more resistant to prompt injection, including against an unseen attacker.","lead":"This paper hardens a web agent against prompt-injection attacks by co-evolving the training tasks, the attacker, and the agent inside a frozen web simulator. The trained 4B model completes more tasks on its own and resists an unseen frontier-model attacker better, and the capability gain carries into a real browser, though the robustness gain is measured only inside the simulator.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The headline Kimi-K3 robustness gain is judged only inside the frozen world model; documented judge errors (C.7) and no browser verification of attacked runs leave the +33.6% gain unverified in executable task completion.","rationale":"The reader's weakest_assumption identifies the same load-bearing premise: all attacked results, including the Kimi-K3 robustness gain, are judge verdicts inside the frozen world model and are not browser-verified, while the judge is documented to accept visibly wrong arithmetic. I find no additional concern that would move the verdict further. The paper is transparent about this limitation (b36, b153, C.7) and the capability claim does have independent browser support (Table 2), which is real evidence. But the 'more robust' claim, especially its strongest out-of-distribution instantiation against Kimi-K3, remains unverified in executable completion. The concrete test—browser-executing the attacked trajectories and scoring with the deterministic contract—would directly test whether the headine gain is a judge/world-model artifact. The reader's CONDITIONAL verdict is appropriate: the paper should be accepted only conditionally on such a verification (or a clearly scoped claim that only simulator-judged robustness is being asserted). No change to the reader's verdict is warranted.","tokens_in":29704,"tokens_out":3362,"duration_ms":31995,"concrete_test":"Execute the Kimi-K3 attacked runs for Base and Robust iter 3 in the real Chromium browser used for Table 2, with the same reactive one-injection-per-trajectory protocol and the same 150 tasks and seeds (0 and 1, and add seed 2 to restore three seeds if feasible). Score each trajectory with the deterministic browser check s(x, tau) defined in Section 3 (strict success: final field values match the contract, protected fields unchanged, no forbidden control activated, review before commit where required) instead of the Qwen judge. Report per-seed strict-success rates for Base and Robust iter 3 under Kimi-K3.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central 'more robust' claim, including the headline Kimi-K3 result (Base 23.00% to Robust iter 3 30.72%, +33.6% relative, Table 3), rests entirely on verdicts of the frozen Qwen3.8-27B judge inside the frozen WebWorld-14B simulator. The paper explicitly states that every attacked result is a verdict of J inside W (b36) and that Stage 2 is not evaluated in the browser (Section 5), acknowledging this limitation (b153). That acknowledgment does not dissolve the concern: it identifies exactly where the load must be borne. For the claim to hold, the judge must grade completion reliably and the world model must render injected content as a real page would. Section C.7 documents the judge labeling visibly wrong arithmetic as positive (10,102 and 8,862 submitted instead of 21,042; six trajectories labeled positive despite totals different from 446.04), showing that the judge can be systematically lenient on content the executor submits. Because the judge is frozen but the executor changes, a trained executor that produces trajectories the judge happens to label more liberally would show an apparent robustness gain that is not a gain in executable completion. The world model can also display values the executor never entered (b153), so the simulator may over-credit successful task execution. The only browser evidence (Table 2) covers capability checkpoints in the clean setting; it does not test any attacked run. Thus the strongest evidence for adversarial robustness—the Kimi-K3 OOD gain—is conditional on unverified judge and world-model fidelity. This is not a finding of error, but it is the single most load-bearing unsecured premise. A concrete browser-based check of the attacked trajectories would settle whether the gain reflects true task-preserving robustness or simulator/judge artifact.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper proposes AdvSim2Real, a two-stage training framework that co-evolves a task-curriculum policy, a prompt-injection adversary, and a web-agent executor inside a frozen web world model (WebWorld-14B), with trajectory completion graded by a frozen LLM judge (Qwen3.8-27B). Stage 1 alternates a curriculum rewarded for tasks the executor solves roughly half the time with executor updates; Stage 2 adds an adversary rewarded only for 'success flips' (a rendered injection that turns a judged clean success into a judged failure) and trains the executor on fresh, historical, and clean attacks. On a released 150-task benchmark, judged clean completion rises from 74.89% to 81.33%, judged completion under the three learned adversaries rises from 48.07% to 57.48%, and judged completion under an unseen frontier adversary (Kimi-K3) rises from 23.00% to 30.72% (+33.6% relative). The Stage-1 capability gain transfers to a real Chromium browser under a deterministic contract check (strict success 25.56% to 44.44%). The paper states explicitly that all attacked results are judge verdicts inside the world model and that Stage 2 was not evaluated in the browser.","tokens_in":30177,"tokens_out":8584,"duration_ms":53262,"significance":"If the judged robustness numbers survive executable verification, the contribution is substantial: the co-evolution of curriculum, adversary, and executor is a novel combination, the success-flip reward is cleanly motivated, and the paper ships code, a benchmark, per-seed counts, and checkpoint results. The browser transfer of the capability stage with a deterministic contract (Table 2), plus the agreement in checkpoint ordering between the world-model judge and the browser check, is genuine evidence that the clean judged gains reflect executed completion. The Kimi-K3 design is the right way to avoid the in-distribution circularity of self-trained adversaries. The documented judge failures (Section C.7) and the absence of any browser verification of attacked runs, however, leave the central 'more robust' claim — and in particular the headline +33.6% Kimi-K3 gain — resting on an unvalidated judge inside a simulator that the paper itself notes can display values the executor never entered.","major_comments":[{"comment":"The Kimi-K3 robustness gain (Base 23.00% vs Robust iter 3 30.72%, Table 3) and every other attacked result is a verdict of the frozen Qwen3.8-27B judge inside the frozen WebWorld-14B simulator, and the paper states Stage 2 is not evaluated in the browser. This is acknowledged (b70, b153), but the acknowledgment identifies exactly where the load is borne: the claim requires the judge to grade completion reliably and the world model to render injections as a real page would. Section C.7 documents the judge labeling visibly wrong arithmetic as positive (10,102 and 8,862 submitted instead of 21,042; six trajectories labeled positive despite totals differing from 446.04), and b153 records that the world model can display values the executor never entered. Because the judge is frozen while the executor changes, a trained executor that produces trajectories the judge labels more liberally would show an apparent robustness gain without a gain in executable completion; Section A.3 quantifies one such route, with 13.4% of verdict flips explained by the final message alone. I recommend verifying attacked outcomes against an executable criterion, for instance by extending the Section C.3 Chromium protocol to replay a sample of attacked trajectories (rendering the injected notice into the served HTML as the world model did) and comparing the judge verdict with the deterministic contract s(x,τ). Without such an audit, the +33.6% Kimi-K3 gain is a judged gain inside one simulator, not a verified gain in executable robustness.","section":"Section 5, Table 3, Section C.7","authors_acknowledge":"Limitations.The world model can display a value the executor never entered and the judge can accept a wrong calculation, so a judged robustness gain is not yet a gain in executable task comple- tion; only the capability checkpoints were verified in a browser (Table 2).","authors_acknowledge_p":0.93},{"comment":"The learned-adversary metrics are in-distribution by construction: Adv v1–v3 are the adversaries that trained the executor during Stage 2, so the 48.07% to 57.48% gain (Table 1) and the 'matches the hosted 9B under attack' comparison (Section 5, against Qwen3.5-9B) are self-play numbers. The authors acknowledge this and correctly introduce Kimi-K3 as the held-out test, but the framing in the abstract, the conclusion, and the hosted-9B comparison still lets the circular numbers carry the robustness claim. I recommend reporting the in-distribution adversary results as self-play diagnostics, resting the cross-adversary claim solely on Kimi-K3, and ideally adding a second held-out adversary trained without any access to the Robust lineage.","section":"Table 1, Section 3, Section 5","authors_acknowledge":"A training procedure that hardens a web agent against injections should be evaluated against adversaries that did not shape its training, as Kimi-K3 does here, and ultimately against one that adapts to the trained agent; blinded human audits, exe- cutable task checkers, and compute-matched component controls are the measurements this requires next.","authors_acknowledge_p":0.61},{"comment":"The judge-accuracy risk is directional and matters specifically for attacked trajectories. The documented false positives (C.7) and the 13.4% of verdict flips attributable to the final message alone (A.3) concern the values and reports the executor submits — exactly the setting where a trained policy might learn judge-pleasing behavior under attack. The paper reports judge-versus-contract agreement only for clean browser runs (Section C.3, Table 6c), not for attacked runs. A concrete test would be to report the agreement between the judge verdict J(x,τ) and the deterministic contract s(x,τ) on attacked trajectories, or a blinded executable audit of a sample of attacked trajectories, for both Base and Robust iter 3. If agreement is high, the concern is resolved; as written, the central robustness claim lacks this check.","section":"Sections A.3, C.7","authors_acknowledge":"Limitations.The world model can display a value the executor never entered and the judge can accept a wrong calculation, so a judged robustness gain is not yet a gain in executable task comple- tion; only the capability checkpoints were verified in a browser (Table 2).","authors_acknowledge_p":0.82}],"minor_comments":[{"comment":"The sentence 'all attacked results are measured inside the world model, where injections are rendered in the context of the page (Section 3)' runs into the next sentence without a period; please fix the punctuation.","section":"Section 1 (page 2)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The conclusion phrase 'matching a hosted 9B model' should carry an explicit qualifier that this comparison uses the in-distribution learned adversaries, not the held-out adversary.","section":"Section 7 (Conclusion)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The caption 'Means over seeds with ±1SD (range over cells in (b))' is unclear about which quantity the 'range' refers to; please state it directly.","section":"Figure 4 caption","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The sentence 'the judge accepts 7.5% to 10.4% of the trajectories that fail it, so judged completion upper-bounds strict completion and ranks the four checkpoints in the same order' links two independent facts with 'so'; upper-bounding does not by itself imply identical ordering, so please separate the two observations.","section":"Section C.3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"In 'its completion rises with and without attacks, holds against a frontier-model adversary it never trained against, and its capability gain carries over to a real browser', the first two clauses are judged metrics while the third is browser-verified; consider making this evidentiary distinction explicit in the abstract.","section":"Abstract","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"The thing to know: this is a real method paper with a genuinely new training loop and an unusually honest limitations section. The headline claim—a 33.6% relative robustness gain against an unseen frontier adversary—is a judge's verdict inside a frozen world model, not an executable outcome. The authors say exactly this themselves (Section 6), so the stress test names the right load-bearing premise, but it is a stated premise, not a hidden one.\n\nWhat's new: the co-evolution of a curriculum that proposes tasks the executor solves about half the time, an adversary rewarded only for success flips, and an executor, all inside a frozen web world model. The success-flip reward (Eq. 5) is a clean idea: replay a judge-accepted clean run to the injection step and credit the adversary only if the judged outcome flips. The combination is not in the ARLAS, CoER, R-Zero, Agent0, GenEnv, WebWorld, or AgentDojo/WASP lines, and the related work is fair, crediting R-Zero and Agent0 for the uncertainty reward. The Stage-1 capability gain is the solid part: strict browser success rises from 25.56% to 44.44%, the checkpoint ordering matches across the judge, the deterministic check, and field counts, and code, benchmark, checkpoints, and per-seed counts are released. That is real evidence.\n\nWhere it is soft. The robustness numbers, including the Kimi-K3 headline, are judge-in-world-model only; Stage 2 is not browser-tested, and the paper says so (Section 5). The documented judge errors in C.7, accepting 10,102 and 8,862 for 21,042, show the judge can be systematically lenient. Since the executor trains against that judge, Stage 2 optimizes judged robustness rather than executable robustness. The clean calibration gap is a concrete hint: 81.33% judged versus 44.44% strict in the browser, though part of that gap is the browser rejecting malformed actions the world model tolerated. I would not call this fatal, because it is acknowledged and addressable, but it is the load-bearing unsecured premise. The learned-adversary numbers (48.07% to 57.48%) are circular against training adversaries; the paper uses them for the hosted-9B comparison, and Kimi-K3 is the only non-circular evidence. Three rollout seeds and no intervals is minor and acknowledged. The ablation does not isolate Stage 1's robustness contribution, and the authors say so themselves; they do not overclaim it.\n\nWho should read it: anyone working on web-agent injection defenses or simulator-based agent training. It deserves a serious referee. A referee should push for browser verification of the attacked runs and a judge that cannot accept visibly wrong arithmetic; with those, this becomes a strong paper.\n\nRecommendation: accept for review, expect revision toward executable verification of the adversarial claims. Engage with it.","headline":"Genuinely new co-evolution training loop with an unusually honest evaluation; the headline robustness gain is a judge's verdict inside a frozen simulator, stated as such by the authors, while the browser-verified evidence covers only the capability stage.","tokens_in":30983,"tokens_out":5094,"duration_ms":53935},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"Co-evolving a task curriculum, an injection adversary, and a web executor inside a frozen web world model makes a 4B agent both more capable and more robust, raising clean completion from 74.89% to 81.33% and completion under an unseen…","keywords":["indirect prompt injection","web agents","adversarial training","co-evolution","web world model","curriculum learning","task-preserving robustness","sim-to-real transfer"],"falsifier":"Run the final 4B checkpoint against Kimi-K3 in the real Chromium browser with the deterministic strict-success check of Table 2, with no world model and no trajectory judge: if strict attacked success does not exceed the base agent's, the claimed robustness gain does not transfer out of the simulator.","tokens_in":31170,"feed_emoji":"🛡️","tokens_out":5616,"duration_ms":33381,"temperature":0.7,"pith_summary":"The paper claims that a web agent can be made both more capable and more resistant to prompt injection by training it entirely inside a frozen web world model against two co-evolving partners: a task curriculum and an injection adversary. The curriculum proposes tasks the agent solves about half the time, pushing capability; the adversary earns reward only for a success flip, an injection that turns a judged success into a failure, pushing robustness. On 150 form-filling tasks, clean completion rises from 74.89% to 81.33%, completion under three learned adversaries from 48.07% to 57.48%, and completion under the unseen frontier-model adversary Kimi-K3 from 23.00% to 30.72%, a 33.6% relative gain. The Stage-1 capability gain carries to a real Chromium browser, where strict success on the submitted form rises from 25.56% to 44.44%, while the attacked robustness gain is only measured inside the simulator.","feed_headline":"Web agent trained in a simulator resists unseen injections","feed_subtitle":"Co-evolving tasks, attacks, and executor lifts attacked completion 33.6% and carries to a real browser.","key_machinery":"The load-bearing mechanism is a pair of rewards feeding LoRA adapters on a 4B executor inside the frozen WebWorld-14B transition model, with a frozen Qwen3.8-27B judge grading every run. The curriculum reward follows R-Zero's uncertainty curve $q(p) = 1 - 2|p - 1/2|$, which peaks at a judged solve rate of one-half and vanishes at rates 0 and 1, so it proposes only tasks near the agent's competence frontier, gated for validity and penalized for repetition. The adversary reward counts only success flips: a clean run the judge accepted is replayed to the injection step, and the adversary earns credit only if the continuation is judged bad, with a rendering check confirming the injection appears. Together they keep both tasks and attacks tracking a moving agent rather than a fixed dataset, and the paper attributes the overall gain to this checkpoint sequence rather than to any single component.","core_discovery":"The paper's central claim is that co-evolving all three policies — task curriculum, injection adversary, and agent executor — inside a single frozen world model makes a 4B executor both more capable and more robust than training on injections fixed before training or on tasks fixed during adversarial training. Stage 1 adapts tasks to the agent's competence frontier and raises clean completion by 4.44 points; Stage 2 adds an adversary rewarded only for success flips and raises attacked completion on the learned adversaries by 3.14 points while clean completion still rises by 2.00 points. The same 4B checkpoint matches a hosted 9B model under attack (57.40% vs 56.96%) and exceeds it clean (81.33% vs 78.22%), and holds 30.72% completion against Kimi-K3, a frontier-model attacker that never saw the training. The paper states plainly that every attacked result is a verdict of the frozen judge inside WebWorld-14B, not a browser-verified outcome, and that only the capability checkpoints were verified in a browser.","pith_inferences":["The 33.6% relative gain against Kimi-K3 could partly reflect the simulator: Kimi's injections render in only 72.3% to 80.9% of episodes and the judge accepts visibly wrong arithmetic as correct, so a browser-based attacked evaluation with the deterministic strict-success check would be the decisive test of whether that gain is real.","Because the adversary reward never verifies that a task remains feasible after injection, training likely produces attacks that remove controls or values the task needs; re-running with an adversary constrained to preserve feasibility would stress-test the defense harder than the current one.","If the uncertainty-curve curriculum were extended to propose attack surfaces as well as tasks, the same co-evolution loop might yield an adversary that adapts to the trained agent, the next test the paper names but does not run."],"forward_implications":["A 4B checkpoint trained on curriculum and adversarial co-evolution matches a hosted 9B model under attack (57.40% vs 56.96%) and beats it clean (81.33% vs 78.22%).","Robustness transfers to an adversary (Kimi-K3) that took no part in training, with completion rising 7.72 points and the gain arriving in the first Stage-2 round.","Training requires no executable site: every transition is predicted by the frozen world model, so tasks and attacks can be generated without hand-built web environments.","Capability gained in Stage 1 transfers to a real browser under a deterministic strict-success check (25.56% to 44.44%), with the four checkpoints ordered identically by simulator judge, browser check, and field count.","Removing Stage 1 costs 4.67 clean points but only 0.44 attacked points, so the ablation does not establish that the curriculum contributes to robustness."],"supporting_citations":[{"why":"WebWorld supplies the frozen transition model that predicts every page, so the entire training and attacked evaluation runs on its simulated pages.","marker":"[Xiao et al., 2026]"},{"why":"R-Zero supplies the uncertainty reward q(p) that the curriculum uses to propose tasks at the competence frontier.","marker":"[Huang et al., 2026]"},{"why":"ARLAS is the co-training attacker-and-agent baseline the paper compares against and distinguishes its alternating-update design from.","marker":"[Wang et al., 2025c]"},{"why":"Documents that adaptive attacks bypass fixed-injection defenses, the motivation for training a learning adversary.","marker":"[Nasr et al., 2026]"},{"why":"WASP quantifies how often indirect injections partially succeed on web-agent benchmarks, setting up the need for task-preserving robustness.","marker":"[Evtimov et al., 2025]"},{"why":"LoRA is the low-rank adapter mechanism through which the curriculum, adversary, and executor policies are each updated.","marker":"[Hu et al., 2022]"},{"why":"WebRL provides the self-evolving curriculum context that Stage 1 adapts, with success estimated from judged completions rather than self-consistency.","marker":"[Qi et al., 2025]"}],"fun_headline_variants":["Co-evolving tasks and attacks makes 4B agent robust in simulator","Simulator training hardens web agent against unseen prompt injections","Adversarial curriculum turns 4B agent into injection-resistant web agent","Adaptive injection training lifts attacked completion by 33.6%"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The whole attacked gain rests on the frozen world model rendering injections the way a real page would and on the frozen judge grading completion correctly, since every attacked result is a simulated verdict; the paper itself records the judge labeling visibly wrong arithmetic as correct.","fun_headline_variants_meta":{"raw":{"variants":["Co-evolving tasks and attacks makes 4B agent robust in simulator","Simulator training hardens web agent against unseen prompt injections","Adversarial curriculum turns 4B agent into injection-resistant web agent","Adaptive injection training lifts attacked completion by 33.6%"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000139,"raw_usage":{"cost":0.000138628,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000138628,"upstream_inference_prompt_cost":0.000109228,"upstream_inference_completions_cost":2.94e-05},"total_tokens":1090,"prompt_tokens":985,"completion_tokens":105,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":30}},"tokens_in":985,"tokens_out":105,"cost_source":"openrouter_usage","duration_ms":2157,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":30,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-07T03:19:41.280094+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Run the final 4B checkpoint against Kimi-K3 in the real Chromium browser with the deterministic strict-success check of Table 2, with no world model and no trajectory judge: if strict attacked success does not exceed the base agent's, the claimed robustness gain does not transfer out of the simulator.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents that adaptive attacks bypass fixed-injection defenses, the motivation for training a learning adversary."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}