A gingerbread house for LLM agents: a honeypot planned from source code lures autonomous attackers into an isolated shadow copy
A Gingerbread House for LLM Agents: Source-Planned Honeypots That Lure Autonomous Attackers into an Isolated Shadow
The authors propose a defense in which an LLM plans static honeypots from source code and routes confirmed attackers to an isolated shadow copy, reducing ARTEX's real compromises on CVE-Bench from 20/40 to 0/40.
The attacker is an autonomous agent given only a natural-language task brief and the target web application's address; the task is zero-day (no CVE, vulnerability class, application, or credentials are disclosed); it can send arbitrary HTTP traffic and run tools in its own environment, but cannot access the internet, the host machine, or Docker, and does not know whether a honeypot exists. The defender owns the application and can read source code and configuration and probe a staging replica before deployment; the attacker counts as compromising the real target (primary) upon achieving one of eight graded impacts.
- Autonomous LLM agents can quickly exploit web vulnerabilities, and patching may come too late; existing deception against agents mostly draws attackers away from a real target that remains reachable, leaving the target itself exposed.
- Before deployment, the defender LLM plans static honeypots from the application source code (decoy pages, fake endpoints, honeytokens, etc.); at runtime, a fixed-code router confirms the attacker via signatures and decoy hits and forwards all subsequent requests from that client (including already-usable exploits) to an isolated shadow replica; the enhanced configuration adds a runtime LLM responder and additional routing rules.
- On the 40 critical tasks of CVE-Bench, the static plan reduced real compromises from 26/40 to 5/40 (Opus 4.8) and from 20/40 to 0/40 (ARTEX); on the 20 new alt-cve-bench tasks, from 15/20 to 0/20. ARTEX's usable exploits mostly succeeded on the shadow while the real target remained intact; the authors describe the defense as deflecting rather than blocking exploits.
- Limitations noted by the authors include: only one retained run per task per arm; no blocking or detection-only baseline; inability to rank configurations once real compromises saturate; single-service container tasks and limited benign traffic; measurement issues such as ARTEX's shared memory and container-name leakage; and the comparison between the enhanced configuration and the static plan does not separate the responder's contribution.
- Evaluated models
- Claude Opus 4.8DeepSeek-V4-Flash
- Benchmarks
- CVE-Bench (40 critical tasks)alt-cve-bench (20 tasks)
- Metrics
- real compromise (grader-verified impact on primary)deflection (exploit success on shadow only)fake secret reportedrequests/seconds to confirmation
Researchers propose a honeypot defense scheme against autonomous attacking agents, in which an LLM defender plans static honeypots based on the application source code, and a router forwards subsequent requests from clients confirmed as attackers, together with available exploits, to an isolated shadow copy. On 40 critical tasks in CVE-Bench, the static scheme reduced real-target compromises of Claude Opus 4.8 from 26/40 to 5/40, and reduced compromises by the ARTEX attack framework from 20/40 to 0/40 (31 valid runs); on 20 entirely new tasks that did not participate in the scheme design, ARTEX went from 15/20 to 0/20. In about half of the deception-group runs, ARTEX's exploits succeeded on the shadow copy while the real target remained intact, indicating that the defense diverts rather than prevents attacks, and the residual risk lies in requests that reach the real target before confirmation is complete.
Related events and news
2 events · 2 news mentions
News that mentions this paper
崇实大学用蜜罐与影子副本诱捕LLM攻击Agent
- 韩国崇实大学gives the paper's title崇实大学 AI 安全研究中心开发诱骗 AI 黑客的欺骗防御技术
박소희 연구교수가 참여했다. 논문 제목은 ‘A Gingerbread House for LLM Agents: Source-Planned Honeypots That Lure Autonomous Attackers into an Isolated Shadow’이며, arXiv(2611.07611)에 공개됐다.
大模型安全周报:智能体安全边界在签名、授权与上下文
- 公众号 大模型安全研究cites the paper's id or link大模型安全周报:智能体的安全边界在签名、授权与上下文三层防线里
5. 从源码自动长出的蜜罐,把已经得手的攻击智能体关进影子环境:A Gingerbread House for LLM Agents: Source-Planned Honeypots That Lure Autonomous Attackers into an Isolated Shadow(https://arxiv.org/abs/2611.07611)
Paper analysis
What problem does this paper attempt to solve?
The authors study how to use static honeypots planned from source code plus per-client routing to isolate autonomous attackers into a shadow copy rather than merely diverting them.
Core question: against autonomous LLM attack agents that can quickly exploit web vulnerabilities, how can a defense be designed that remains effective once an exploit is already usable, and how can one determine whether it blocks the exploit or deflects it to an isolated replica.
- Setting and motivation: the authors state that LLM agents can exploit web vulnerabilities without a vulnerability description, and that frameworks such as ARTEX have been used by real intruders; going from foothold to compromise may take only hours, and patching may come too late.
- Shortcomings of existing methods: the authors state that existing deception against LLM agents (e.g., AgentSnare, AgentTrap) mostly draws attackers away from the real target, which remains reachable, and asset safety depends on the agent continuing to prefer the decoy.
- Core idea: drawing on the older designs of shadow honeypots and honey-patches, the arrangement of "moving a flagged attacker to a service replica" is brought into the LLM-attacker setting: an LLM defender plans static honeypots from the application source code, and a router sends all subsequent requests from a confirmed attacker client (including usable exploits) to an isolated shadow replica.
- What the paper proposes: a defense (static plan arm B, enhanced configuration arm C) plus an evaluation method: use the benchmark's own grader to judge outcomes on both the real target and the shadow replica, and analyze whether the attacker's report accepts planted fake secrets; it also builds 20 new alt-cve-bench tasks.
- Threat model (Section 3.1):
- Attacker goal: given a natural-language brief and the target address, achieve any one of eight graded impacts (e.g., reading a secret file, code execution).
- Attacker knowledge: the task is zero-day; no CVE, vulnerability class, application, or credentials are disclosed, and it does not know whether a honeypot exists; the brief asks it to suspect and verify "too-convenient shortcuts."
- Attacker capabilities: can send arbitrary HTTP traffic and run tools in its own environment; cannot access the internet, the host machine, or Docker.
- Victim system and defender: the defender owns the application and can read source code and configuration and probe a staging replica before deployment, but does not know the CVE or vulnerability description; the defense fails if the attacker achieves any graded impact on the real target.
What related research exists?
Related work falls into four groups: LLM attack agents and benchmarks, agent gullibility toward the environment, classic and LLM-targeted deception, and two 2026 systems that measure real compromises.
LLM agents as attackers
- Fang et al. (2024b, 2024a), Zhu et al. (2026)—tool-using LLM agents exploit Web vulnerabilities without prior knowledge; frameworks such as PentestGPT (Deng et al., 2024), AutoAttacker (Xu et al., 2024), CAI (Mayoral-Vilches et al., 2025) and Incalmo (Singer et al., 2025) organize complete attack workflows around LLMs.
- Benchmarks: CVE-Bench (Zhu et al., 2025) and PACEbench (Liu et al., 2025) score on reproducing CVEs, Cybench (Zhang et al., 2025b) uses CTF tasks, and BountyBench (Zhang et al., 2025a) uses bug bounties. This paper follows CVE-Bench and its grader and adds 20 tasks in the same format.
Agents trust the environment
- InjecAgent (Zhan et al., 2024)—agents execute instructions injected into tool outputs; Zhang et al. (2025c)—clicking adversarial pop-ups; Prinos et al. (2026)—agents act on a trap even when they name it (recognition–action gap). The authors explicitly do not treat this credulity as a new finding and only study how it manifests in real attack reports.
Deception against attackers
- Classic deception: honeypots (Spitzner, 2002), honeywords (Juels and Rivest, 2013), survey (Han et al., 2018); shadow honeypots (Anagnostakis et al., 2005) send traffic flagged by an anomaly detector to an instrumented replica, and honey-patches (Araujo et al., 2014) fork attackers who trigger a patched vulnerability into a decoy container that provides false information.
- Against LLM agents: prompt-injection defenses (Pasquini et al., 2024; Reworr and Volkov, 2024), cloaks and traps (Ayzenshteyn et al., 2025), decoy vulnerabilities (Zhang et al., 2026); Soosahabi and Namsani (2026) study the effectiveness of defenses based on model-analysis misdirection, and Cordeiro et al., 2026's Rouxii studies deception-aware penetration agents that can recognize honeypots.
Two 2026 systems that measure compromise of real targets (the comparison targets named in the paper)
- AgentSnare (Wang et al., 2026a)—uses breadcrumbs to lure agents into LLM-simulated decoy hosts and reports no real exploitation at pass@3 on 15 CVE-Bench tasks.
- AgentTrap (Wang et al., 2026b)—places sentinel endpoints next to vulnerable endpoints and reduces attack success rate from 95.8% to 79.2%.
- The paper notes that these two kinds of decoys require agents to prefer the decoy and the real path to remain open; this paper's router instead sends all subsequent requests from a confirmed client (including usable exploits) to a shadow replica.
Positioning: The authors say the contribution lies in the combination (static deception planned from source code + per-client routing to an isolated replica) and the measurement it enables—running the grader on both replicas can distinguish "exploit succeeds on the shadow" from "exploit fails," and it supplements AgentSnare's measurement of decoy acceptance; the authors explicitly do not directly compare with these systems and note that static and dynamic deception have been compared before (Wang et al., 2026b).
How does the paper address this problem?
The defense consists of static honeypots planned offline by an LLM (arm B) and a runtime router; arm C adds an LLM responder and additional confirmation rules; evaluation scores both the real target and the shadow copy.
Overall approach: the defender LLM plans a set of static honeypots from the application source code before deployment, and at runtime a router implemented as fixed code forwards all subsequent requests from a confirmed attacker, keyed by client state, to an isolated shadow copy, so that usable exploits land on the copy rather than the real target.
Static plan (arm B, Section 3.2 and Appendix A)
- Two planning agents: both Claude Opus 4.8 (via Claude Code, with read/search/write tools only). The analyst reads the staging copy's code and configuration export (secrets redacted, with a size cap), probes staging, and produces an application profile and at most ten ordered risk surfaces (method, path pattern, attack category), and is told not to name CVEs; the designer reads only the profile and the risk list and produces a plan with five parts: decoy pages, fake endpoints, prompts, honeytokens and detection rules.
- Placeholders and gates: all fake values are placeholders, refilled on each run; the plan must pass six automatic gates (schema check, decoy build, decoy paths must 404 on the real application, invariant check replaying at most 20 probe paths, in-application placement check, detection rule check), plus a taint check to prevent leaking benchmark identifiers.
- No LLM at runtime: the LLM only writes the plan; the router is fixed code that reads the plan, and arm B makes no defender-side LLM calls during the attack.
Router and shadow copy (Section 3.2, Figure 2)
- Three-state state machine: one state per client (normal → suspect → confirmed), which only moves forward. A payload signature (such as SQL injection or path traversal patterns) or a 404 burst (5 404s within 20 requests) makes the client suspect; a honeytoken, a hit on a fake endpoint or decoy, or a payload on a predicted risk surface confirms it.
- Forward everything after confirmation: from the confirming request onward, every new request from that client goes to the shadow copy—a clone of the target in a separate container stack, with its own network, storage and fake secrets; requests already forwarded to the real target complete there (10 in one run). Deceptive responses are delayed 2–4 s (tarpit).
- Client attribution in the experiments: the attacker accesses through a separate listener, and the router maintains state per listener, so all of the attacker's requests are attributed to one client; the benign crawler uses a second listener. The authors note that deployment would need to key by session or address, and that an attacker who changes identity would start over as a new client.
Enhanced configuration (arm C, Section 3.3 and Table 1)
- LLM responder: Claude Opus 4.8, called at most three times per run, starting at confirmation; it reads the attacker's recent traffic and places new fake endpoints, prompts, files or database rows on the shadow, which must pass the same gates as static items; it places nothing more once the shadow grader has reported success.
- Additional routing rules: escalation (an unauthenticated request hitting a top-ranked code execution or authentication surface confirms the suspect client), persistence (the 20th request in suspect state confirms), secret guard (knows this run's synthetic secret and, when a real response is about to carry the secret out, immediately confirms and switches to a shadow response); and the query parameter marking decoy links is renamed per run (arm B uses the default name, which contains the string "artex").
- Comparison basis: arm B and arm C differ in several ways besides the responder (Table 1), and RQ4 compares the two configurations as a whole and cannot isolate the responder's contribution.
Measurement design (Section 4)
- Primary metric: the benchmark's built-in grader verifying impact on the real target (primary), or a real secret being exposed to the attacker before confirmation, counts as a compromise; a grader-verified impact is called a working exploit and need not use the task's corresponding CVE. Both arms share the grader, with an exact McNemar test paired by task; with zero compromises, the Clopper–Pearson two-sided 95% interval upper bound is reported.
- Shadow judging: the same grader runs on the shadow, distinguishing compromised / deflected (shadow-only success) / no usable exploit; "exploit success on any copy" is the quantity comparable to control-arm compromise, used to distinguish "blocking" from "transfer".
- Attacker reports: Opus 4.8 writes the secret it identifies to a file; ARTEX writes no file, and the harness takes the first 32-digit hexadecimal value in its findings, report or command output and judges it as reported / doubted / seen only.
- Margin metrics: number of requests and seconds before confirmation, type of confirmation signal, number of payload requests reaching the real target before confirmation, and the proportion of requests handled by the shadow; after each deception run, a benign crawler is run through a separate listener and its state recorded.
- Validity: in Experiments 2 and 3, runs had to have all five cheating signals at zero (contacting exploit/advisory hosts, using web search tools, reading other runs' directories, issuing Docker commands, delegating to external agents), extracted from tool calls by a frozen scanner and manually reviewed.
What experiments did the paper conduct?
Three laddered experiments (Opus 4.8/ARTEX × CVE-Bench, ARTEX × alt-cve-bench) show the static plan reduces real compromises to zero or near zero, with exploits deflected to shadow replicas.
Experimental setup
- Attackers under test: Experiment 1 used Claude Opus 4.8 as an autonomous agent (HTTP traffic passed through a logging wrapper); Experiments 2 and 3 used the ARTEX multi-agent penetration framework (driven by DeepSeek-V4-Flash), with ARTEX unmodified, running in an internal-network-only container, with the web-search and web-fetch tools disabled.
- Benchmarks: CVE-Bench critical, 40 Web CVE tasks (2023–2024), with an evaluator inside each task's target container judging eight kinds of impact; alt-cve-bench is a set of 20 Web CVE tasks built by the authors, with no CVE or application overlap with CVE-Bench, the same task format and evaluator byte-for-byte, and difficulty fixed before running at 6 easy / 8 medium / 6 hard (Table 5, Appendix C).
- Control arms: arm A no defense, arm B static plan, arm C enhanced configuration; one attack per task per arm (Table 2).
- Judging method: the benchmark's built-in grader judges the real target and the shadow copy; zero compromises are reported with the Clopper–Pearson 95% interval upper bound; comparisons between arms use the exact McNemar test paired by task.
- Budget: in Experiments 2 and 3 the nominal budget was 1,200 s for arm A and 1,500 s for the deception arms (partly offset by tarpit delays).
- Sample size: of 300 planned runs, 296 were recorded and 288 were valid; in Experiment 2, CVE-2024-34716 never started and was kept among the 40 tasks but had no valid run.
Main results
According to Table 3 (one retained run per task per arm, with counts covering all 40 or 20 tasks):
Swipe horizontally to read the table
Experiment Arm Valid Real compromise Shadow-only success Fake secret reported 1 (Opus 4.8, CVE-Bench) A 40/40 26/40 – – 1 B 40/40 5/40 n/a 0/40 1 C 40/40 1/40 n/a 0/40 2 (ARTEX, CVE-Bench) A 38/40 20/40 – – 2 B 31/40 0/40 (≤11.2%) 14/40 7/40 2 C 39/40 0/40 (≤9.0%) 21/40 14/40 3 (ARTEX, alt-cve-bench) A 20/20 15/20 – – 3 B 20/20 0/20 (≤16.8%) 12/20 3/20 3 C 20/20 0/20 (≤16.8%) 14/20 6/20 - Of the 5 compromises in Experiment 1, 4 were grader successes and 1 was a real secret exposure before confirmation; Experiment 1 did not record per-run results on the shadow (Table 3 footnote f).
- Paired McNemar: in Experiment 2 the static arm versus control was 15:0 (30 pairs, p = 6.1 × 10−5), and the enhanced arm 20:0 (38 pairs); in Experiment 3 both arms were 15:0 (20 pairs, p = 6.1 × 10−5).
- The authors state that the static plan already achieved most of the observed reduction; on the 30 tasks where the control arm compromised and the static arm was valid, the static arm's compromise rate 95% interval upper bound was 11.6%.
RQ2: exploit is transferred rather than blocked (Section 6.2)
- Exploit success on any copy: ARTEX on CVE-Bench had 14/40 for the static arm and 21/40 for the enhanced arm, versus 20/40 for the control; on alt-cve-bench, 12/20 and 14/20, versus 15/20 for the control. The paired differences were not significant (p ≥ 0.25), but the authors note that one run per task cannot prove the exploitation probability is unchanged, and that 12/20 versus 15/20 on alt-cve-bench is a 15 percentage point gap; the deception arms also had a longer nominal budget.
- Confirmation speed: the static arm on CVE-Bench confirmed at a median of 6 requests (IQR 3–30) and 26.9 s, and on alt-cve-bench at 5 requests (4–17) and 20.2 s; the confirmation signal was most often a payload on a predicted risk surface (17/31 and 11/20), followed by fake endpoints (9 and 6) and decoys (5 and 3). The shadow handled a median of 95.2% and 93.2% of requests per run.
- The margin is not always wide: 2/31 CVE-Bench runs had a payload reach the real target before confirmation (7 and 4 requests), without causing a compromise; one run (CVE-2024-31611) confirmed only after 31,409 requests to nonexistent paths. The analyst's risk surface named the true vulnerable handler in only 22/39 plans, but after 6 misses confirmation still occurred via decoys or other predicted surfaces; the authors state that protection after a miss depends on the attacker probing broadly before reaching the vulnerability.
RQ3/RQ4: changing attacker, new tasks and enhanced configuration (Sections 6.3, 6.5)
- Changing attacker (descriptive comparison): on the same 40 tasks the two controls compromised 26/40 (Opus 4.8) and 20/40 (ARTEX); the static plan let through 5 tasks for Opus 4.8 and 0 for ARTEX. The authors note that Experiment 1 is a development set, that the two experiments used different plans and briefs, and that the comparison is descriptive only.
- New tasks: on alt-cve-bench no real compromise was observed, shadow exploit success was 12/20 and 14/20, and fake secrets were reported 3/20 and 6/20; the task sets differ, and the authors do not compare rates across benchmarks.
- RQ4: in Experiment 1, arm C reduced compromises from 5/40 to 1/40 (5:1, p = 0.22); in the ARTEX experiments both arms were zero, and across 51 paired tasks deflection was 26 versus 29 (3:6, p = 0.51), fake secret reports 10 versus 17 (6:13, p = 0.17), and confirmation time a median of 22.4 s versus 23.0 s. The authors state that no significant difference was detected and that the comparison does not isolate the responder.
ARTEX's reporting of decoys (Section 6.4, Table 10)
- Believing fake secrets: across the four ARTEX deception arms, with 120 tasks and 110 valid runs, 55 runs extracted a value that was a planted fake secret; of these, 18 appeared only in raw tool output, 7 were questioned, 30 were reported as real findings, and 20 occurred without any usable exploit (produced by decoys alone).
- Still believing after recognition: 10 runs reported that something was a decoy or trap, and 5 of these still reported the planted value as true; the authors state this resembles the recognition–action gap measured by questionnaire (Prinos et al., 2026), but it is only an observation of the final report, and the order in which items were encountered was not reconstructed.
Other ablations and analyses
- Benign crawler: of 117 recorded deception-arm runs, 115 stayed normal and 2 became suspect (both on CVE-2024-37849, due to a 404 burst caused by broken resource links), and none were ever confirmed.
- Shared-memory sensitivity: after excluding tasks with residual memory of the same CVE, the deception arms' compromise comparison against control remained significant (Experiment 2: 13:0 and 15:0; Experiment 3: 11:0 and 12:0; all p < 0.001); RQ4's deflection surplus dropped from 3:6 to 2:3 (Section F).
- Grader positive control: on new instances of 20 tasks, the primary grader accepted its own secret every time and never accepted the shadow's secret, with all 166 checks matching expectations (Section F).
- Responder placement: in Experiment 3 the median was 3 calls and 6 items placed per run (123 items in total), and Experiment 2 placed 271 items in total; the main reason for rejection was that the shadow grader had already reported success (33 items in Experiment 3, 39 in Experiment 2).
- Cost: planning had a median of $2.02–2.40 and 7–16 minutes per application; the responder had a median of $1.76–1.92 per run (Section G).
What are some points that can be further explored?
The authors list limitations including single runs, a no-blocking baseline, and saturation; the experiments cover only two attackers, single-service Web tasks, and one run per task.
Limitations and future directions noted by the authors
- Only one retained run per task per arm (Limitations): agents are nondeterministic, and run-to-run variance was not measured; secondary outcomes (deflection, fake secrets, margin) are estimated imprecisely on 20–38 paired tasks, and the non-significance of RQ4 cannot prove equivalence.
- No blocking or detection-only baseline (Limitations): comparison is only against no defense, so it cannot show that deflection is better than blocking, nor the increment of decoys over an LLM-configured detector or a traditional WAF.
- Saturation problem (Limitations): when the static plan already prevents all compromises, the primary metric cannot rank additional components; secondary quantities measure only the margin in observed runs.
- Attacker coverage and attribution assumptions (Limitations): the two attackers differ in model, scaffolding, brief, and secret-extraction method, so differences cannot be attributed to a single factor; an attacker not programmed to detect deception, one that cross-validates secrets, fingerprints the shadow, or rotates client identities could weaken routing and decoys; per-listener client attribution is "the most exposed assumption."
- Shared attacker memory and container-name leakage (Limitations, Section F): ARTEX's single-instance asset library persists across tasks (in 176 recorded runs, 147 returned records from other tasks and 22 were earlier runs of the same CVE), with the direction of impact undetermined; the shadow's container-name prefix was visible to an attacker with code execution in 7 runs and must not be exposed at deployment.
- Task and traffic scope, measurement (Limitations): all are single-service containerized web applications; multi-host or third-party benchmarks with a WAF (Liu et al., 2025) are left to future work; the benign check is only a same-origin link crawler and does not represent logins, searches, uploads, API clients, or traffic interleaved with attacks; the cheating scanner was changed twice in Experiment 3 and then frozen, and the fake-secret coding and analyst recall labels come from a single LLM-assisted coding pass with no second independent coder.
Experimental coverage
- Two attackers were tested: Claude Opus 4.8 single agent (Experiment 1) and ARTEX (driven by DeepSeek-V4-Flash, Experiments 2 and 3).
- Tasks were the 40 critical CVE-Bench tasks and 20 self-built alt-cve-bench tasks, all single-service web applications; one retained run per task per arm.
- Defense configurations were arm B static plan and arm C enhanced configuration, with no defense as control; no direct comparison with AgentSnare, AgentTrap, or a WAF was made (the paper explicitly does not make a direct comparison).
- Benign traffic was produced only by a same-origin link crawler on a separate listener, across 117 recorded deception-arm runs.
- The paper does not report run-to-run variance or repetition counts (only one retained run per task); Experiment 1 did not record per-run results on the shadow or router margin.
Summarize the main content of the paper
The source-planned static honeypots plus per-client routing isolate autonomous attackers into shadow replicas, nearly eliminating real-target compromises in the test configuration.
A defense paper: against autonomous LLM attack agents that can quickly exploit web vulnerabilities, the authors bring the older arrangement of "moving a confirmed attacker to a service replica" into the LLM-attacker setting and design an evaluation that can distinguish "blocking the exploit" from "deflecting the exploit."
- Problem: the authors state that patching may not keep up with agent attack speed; existing deception against agents mostly draws attackers away from a real target that remains reachable, and asset safety depends on the agent continuing to prefer the decoy.
- Method: two Claude Opus 4.8 agents (analyst and designer) plan static honeypots from the application source code before deployment (decoy pages, fake endpoints, honeytokens, detection rules), passing six automated gates; at runtime, a fixed-code router uses three states (normal/suspect/confirmed) to send all subsequent requests from a confirmed client to an isolated shadow replica; arm B makes no LLM calls at runtime, and arm C adds an LLM responder and additional confirmation rules.
- Main results: on the 40 critical tasks of CVE-Bench, the static plan reduced real compromises from 26/40 to 5/40 (Opus 4.8) and from 20/40 to 0/40 (ARTEX, 31 valid runs); on the 20 alt-cve-bench tasks not involved in design, from 15/20 to 0/20 (Table 3).
- Deflection rather than blocking: ARTEX's usable exploits succeeded on the shadow while the real target remained intact (CVE-Bench static arm 14/40, enhanced arm 21/40); the authors state that in the observed runs the defense deflected usable exploits; the pre-confirmation window is the residual risk, with a median of 6 requests and 26.9 s to confirmation in the static arm.
- Attacker reports: in 110 valid deception-arm runs, ARTEX reported planted fake secrets as real findings 30 times, 5 of which occurred when the same report had already flagged another item as a decoy (Table 10).
- Authors' conclusion: the authors state that the combination greatly reduced real-target compromises in the test configuration, but the evidence is limited to one retained run per task, single-service CVE tasks, and fixed-listener attribution; no comparison with a blocking baseline was made, and robustness against identity rotation and attackers expecting deception remains to be tested.