Optional Traps: An Empirical Study of LLM Agents Facing Credential Deception and Alert-Driven Revocation
Optional Traps: An Empirical Study of LLM Agents Against Credential Deception and Alert-Driven Revocation
The authors built a honeypot experimental platform to test whether LLM agents touch decoys: the 9.7B model bypassed all optional decoys and directly exploited the real vulnerability; the 4B model across 60 sessions showed no additional defensive effect from deception after hardening.
- Traditional deception defenses (honeypots, decoy credentials) assume that legitimate users will not touch fake assets, but whether LLM agents are subject to this constraint, and whether deception still has incremental value after conventional hardening, still lacks empirical testing.
- Build an isolated hash-chained log laboratory containing a simulated Records API and OpenCanary sensors, and through four studies compare baseline, hardened-only, and hardened+deception conditions, testing LLM agents' touching of decoys and the alert-driven revocation mechanism.
- The 9.7B model ignored all optional decoys and directly exploited the real vulnerability; the 4B model across 60 sessions showed hardening alone completely prevented export (0/20), with deception only adding detection and revocation (19/20) and no additional prevention; cloud validation completed revocation within 303ms without affecting legitimate clients.
- Small sample size (single target, 4B/9.7B local models); Study 2 was not externally pre-registered; the hash chain is not signed; IP-based identity may fail under shared NAT; larger or commercial models were not tested.
- Threat model
- The attacker is an LLM agent with path-restricted HTTP tools and a submission tool, can read arbitrary responses and use discovered credentials, but cannot execute code on the target, access the internet, or tamper with sensors; the defender controls the gateway, sensors, and adapters, and client identity is based on the socket peer or HMAC-signed headers.
- Evaluated models
- tantular-cyber:latestqwen3.5:4b-q8_0
- Benchmarks
- Mock Records APIOpenCanary
- Metrics
- Genuine export successPeer revokedExact paired McNemar pWilson 95% CI
The researchers test in an isolated experimental environment whether deception measures such as honeypots and canary credentials can defend against LLM agents stealing credentials. In 60-session experiments with a 4B model, baseline exfiltration of protected records succeeded 20/20, while hardening alone and hardening plus deception were both 0/20; deception produced detection and per-client revocation in 19 of the 20 defended sessions, but provided no measurable preventive gain beyond hardening. On Google Cloud, a real OpenCanary login alert triggered revocation of the attacking client, with the first denied exfiltration appearing within the observed upper bound of 303 ms, while another legitimate client's 100/100 requests were unaffected. In multi-agent experiments, an agent that touched the canary was revoked 39-59 seconds after the decoy was deployed without affecting its peers, but one model-driven legitimate agent also touched the decoy by mistake, constituting a false-positive risk specific to LLM clients; in exploratory runs, a 9.7B model also ignored all optional decoys and used a genuinely leaked service key to exfiltrate records.
Paper analysis
What problem does this paper attempt to solve?
Test whether LLM agents touch optional decoys, and whether deception still has incremental defensive value after routine hardening.
Core question: traditional deception defenses (honeypots, decoy credentials) rely on the assumption that "legitimate users never touch fake assets", but whether this also holds for LLM agents, and whether deception provides incremental value on top of conventional hardening measures, still lacks empirical study.
- Problem setting and importance: LLM agents can now autonomously chain reconnaissance, credential discovery, and exploitation to attack web targets. Defenders have long relied on deception techniques (honeypots, honeytokens, honeywords) to detect such activity early, but the authors believe it needs to be verified whether these techniques remain effective against LLM agents.
- Shortcomings of existing methods: existing deception research mostly targets human attackers, or advocates using prompt injection to counter LLM agents. The authors point out that there is a lack of controlled experiments comparing "with and without deception" under the same hardening conditions, and that prompt injection relies on model-specific behavior and is not general enough.
- Core observation and hypothesis: the authors hypothesize that LLM agents may completely bypass optional decoys placed beside the attack path and directly exploit the real vulnerability; the main value of deception may lie in detection and attribution rather than prevention.
- Threat model:
- Attacker goal: use an LLM agent to discover credentials and export protected records through path-restricted HTTP tools.
- Attacker knowledge: can read arbitrary responses and use any discovered credentials, but cannot execute code, access the internet, or tamper with sensors.
- Defender capabilities: controls the gateway, sensors, and adapters, can identify clients based on the socket peer or HMAC-signed headers, and can carry out revocation.
- Victim system: a simulated Records API containing a deliberately planted debug endpoint vulnerability (V1 leaks the real key; V2 returns 404 or plants a canary credential).
What related research exists?
Existing work covers honeypot techniques against human attackers, LLM agents' autonomous attack capabilities, and prompt injection defenses against LLMs.
Deception techniques against human attackers
- Spitzner (2002), Stoll (1989): Proposed the concept of honeypots, exposing systems with no production purpose, where any interaction is suspicious.
- Juels and Rivest (2013): Proposed honeywords, planting decoy passwords in password files so that use immediately exposes cracking activity.
- MITRE Engage (2022): Framed denial and deception as planned adversarial engagement with explicit objectives.
LLM agents as attackers
- Fang et al. (2024): Demonstrated that tool-using LLM agents can autonomously exploit web application vulnerabilities, including multi-step attacks, without prior knowledge of specific flaws. This motivates defenses to assume attackers will read and reason about server responses.
Deceptive defenses against LLM agents
- Pasquini et al. (2024): Proposed Mantis, embedding prompt injection payloads in decoy services to disrupt or counter LLM-driven attackers.
- Ayzenshteyn et al. (2025): Proposed CHeaT, an active defense against LLM penetration-testing agents implementing asset masquerading, honeypot placement and trap setting.
- Reworr and Volkov (2024): Deployed internet-facing honeypots embedding prompt injection and timing analysis to detect AI hacker agents in the wild.
Baselines and comparisons
- Baseline methods: The paper did not use external baseline methods, but instead set internal control conditions: baseline (no hardening, no deception), hardening (vulnerabilities closed, no deception), defense (hardening + deception).
- Benchmarks used: No public benchmark; used a self-built Mock Records API and OpenCanary honeypot software.
The authors position this paper's difference from the above work as: avoiding prompt injection and hidden instructions, using only ordinary fake assets and canary credentials, making the defense independent of model-specific behavior; the core methodological contribution is the hardening-comparison-defense ablation experiment in Study 2, separating the preventive effect of removing real vulnerabilities from the preventive effect of deception.
How does the paper address this problem?
Build an isolated hash-chained log laboratory, and through four studies compare baseline, hardened, and hardened+deception conditions, testing LLM agents' touching of decoys and alert-driven revocation.
The overall idea is to build an isolated, reproducible deception laboratory, design four studies around a simulated Records API, and by comparing baseline, hardened-only, and hardened+deception conditions, separate the defensive effects of hardening and deception and validate the alert-driven revocation mechanism.
Laboratory architecture and threat model
- Target system: the Mock Records API provides synthetic records, and the real goal is to call GET /api/v2/records/export with a valid bearer key; the deliberate vulnerability is the debug endpoint /api/v1/debug/env (V1 and the V2 baseline leak the real service key; after hardening, V2 returns 404 or plants a canary key).
- V1 deception modes: static mode adds a backup file containing a canary key and a deprecated internal API listed in robots.txt; adaptive mode makes the internal API return endless pagination leads. Using the canary key or 12 decoy hits terminates the session within the API.
- V2 defense architecture: a separate gateway (Figure 1) replaces the vulnerability under three conditions: baseline (leaks the real key), hardened (debug route returns 404), and defense (debug route returns a planted canary key and enables enforcement). Canary use revokes the triggering peer for 15 minutes.
- Threat model: the attacker is an LLM agent with path-restricted HTTP tools and a submission tool, can read arbitrary responses and use discovered credentials, but cannot execute code on the target, access the internet, or tamper with sensors; the defender controls the gateway, sensors, and adapters, and client identity is based on the socket peer (HMAC-signed headers in Study 4), ignoring X-Forwarded-For.
Evidence and isolation mechanisms
- Budgets and limits: agents run under hard budgets (HTTP calls, model turns, report tokens, wall-clock time), with tokens taken from model runtime reports rather than estimates.
- Hash-chained logs: every agent and defense event is appended to a SHA-256 hash chain; the verifier recomputes digests, coverage, aggregate counts, and statistics. The logs are not signed and can detect inconsistency but do not prove authorship.
- Isolated environment: containers run on an internal Docker network with no published ports, non-root users, read-only filesystems, no capabilities, and resource limits; the local container laboratory passes 60/60 isolation and enforcement checks.
Study design
- Study 1 (exploratory): uses tantular-cyber:latest (9.7B parameters, Q8_0, local Ollama), seed 13112026, one session per V1 mode, against a fresh loopback instance of the Records API; budget is 25 HTTP calls, 30 model turns, 60,000 report tokens, 180 seconds wall-clock time.
- Study 2 (controlled experiment): uses qwen3.5:4b-q8_0 (local Ollama), temperature 0.2, 8,192-token context, 256 tokens generated per turn, thinking disabled; three conditions (baseline, hardened, defense) with 20 trials each, 60 sessions in total, condition order rotated, shared sampling seeds (starting from 131200) and the same task; each session uses a new target and a new conversation, with randomly synthesized credentials; budget is 8 HTTP calls, 8 model turns, 30,000 report tokens, 90 seconds wall-clock time; the task tells the agent the discovery and export routes, measuring exploitation of a known vulnerability rather than route discovery.
- Study 3 (cloud validation): runs a separate V2 stack on a Google Cloud e2-small VM (Ubuntu 24.04), accessed via Identity-Aware Proxy SSH; the runner generates a new synthetic key, executes 14 checks in a fixed order, and removes the stack; the verifier deliberately gives the attacker container a valid laboratory token, testing detection rather than token-based access decisions.
- Study 4 (multi-agent swarm): a swarm of role-prompted agents (reconnaissance, collection, exploitation, decoy) and one legitimate agent attack the V2 gateway on a GCP VM through an IAP tunnel; each agent carries an HMAC-signed peer identity, and revoking one does not affect the others; models run via local Ollama on the operator machine.
Statistics and verification
- Primary outcome: paired true-target success rate, baseline vs defense, using an exact two-sided McNemar test; the prespecified threshold requires at least 20 valid pairs, a baseline success rate of at least 80%, fewer successes under defense, p < 0.05, and at most 10% invalid runs per condition.
- Ablation test: hardened vs defense is a secondary ablation, and no incremental benefit is claimed unless independently supported; interval estimates use Wilson 95%.
- Independent verification: the verifier recomputes the results and protocol digests, source fingerprints, session coverage, aggregate counts, Wilson intervals, and exact paired statistics.
What experiments did the paper conduct?
Four studies: the 9.7B model bypassed all optional decoys; the 4B model across 60 sessions showed hardening alone was sufficient to prevent export, with deception only adding detection; revocation completed in 303ms in the cloud; a legitimate agent in the swarm mistakenly triggered a decoy.
Experimental setup
- Models tested: tantular-cyber:latest (9.7B parameters, Q8_0, local Ollama, Study 1); qwen3.5:4b-q8_0 (4B parameters, local Ollama, Studies 2 and 4).
- Datasets/benchmarks: No public benchmark; used a self-built Mock Records API (synthetic records and credentials) and OpenCanary honeypot software.
- Baselines: Internal control conditions—baseline (no hardening, no deception), hardening (vulnerabilities closed, no deception), defense (hardening + deception).
- Metric definitions: Genuine export success, peer revoked, exact paired McNemar p-value, Wilson 95% confidence interval.
- Review method: No LLM review; independent verifiers recomputed hash chains, statistics and protocol compliance.
- Sample size and repetitions: Study 1: 1 session per mode; Study 2: 20 trials per condition, 60 sessions total; Study 3: single run of 14 checks; Study 4: four independent engagements (A, B, C, D).
Main results
Swipe horizontally to read the table
Condition Debug route behavior Genuine export success Wilson 95% CI Peer revoked Baseline Leaked real lab key 20/20 (100%) 83.9%–100% 0/20 Hardening Returned 404 0/20 (0%) 0%–16.1% 0/20 Defense Provided planted canary key, enabled execution 0/20 (0%) 0%–16.1% 19/20 According to Table II. Exact paired McNemar test: baseline vs defense p = 1.9073486328125 × 10−6; hardening vs defense p = 1 (no discordant pairs).
- Authors' interpretation: The hardening and defense conditions are indistinguishable in export success rate (both 0/20), indicating the preventive effect comes entirely from closing the real vulnerability; the incremental value of deception lies in detection and attribution (19/20 sessions triggered revocation), not additional prevention.
- Exception: In trial 11 (seed 131210), the model read the decoy route but submitted unrecognized credentials to the export endpoint (HTTP 401), then queried nonexistent debug routes until the budget was exhausted, neither exporting nor triggering revocation.
Study 1: Live agent bypasses optional decoys
- Setup: tantular-cyber:latest (9.7B), 1 session per mode, budget of 25 HTTP calls, 30 model turns, 60,000 tokens, 180 seconds.
- Results: Under both static and adaptive deception modes, the model ignored all decoys, read the real key from /api/v1/debug/env within 6 HTTP calls and called /api/v2/records/export; in the baseline session the model also found the key but kept trying the wrong /api/v1/... export path until the wall-clock budget was exhausted (Table I).
- Authors' interpretation: When the real vulnerability is reachable, decoys placed alongside the path are simply skipped; the baseline failure is a capability deficit rather than a defensive effect.
Study 3: Cloud alert-driven revocation validation
- Setup: Google Cloud e2-small VM, OpenCanary sensor, HMAC-signed adapter, 14 control checks.
- Results: A genuine OpenCanary login alert caused the attacker client's valid token to be rejected within 303 ms (observation upper bound); legitimate clients completed 100/100 requests; all 14 checks passed (Table III), including forged sensor webhook returning 403, replayed V1 debug key bypass returning 403 and revoking, invalid HMAC signature returning 401, etc.
- Authors' interpretation: Commodity-grade honeypot alerts can be authenticated, edited, signed and turned into per-client revocation in hundreds of milliseconds without disturbing coexisting legitimate clients.
Study 4: Multi-agent swarm engagement
- Setup: Four engagements (A: script-driven, no model; B: qwen3.5:4b-q8_0; C, D: tantular-cyber), attacking the V2 gateway through an IAP tunnel.
- Results: Agents using the canary were revoked 39–59 seconds after the decoy was offered (Table IV), with 58 subsequent requests denied; in Engagement B a legitimate LLM agent also touched the decoy and was revoked (false positive); in Engagement C, 6 agents shared a local runtime, averaging only 5 model turns within 180 seconds, with containment coming from the budget rather than the defense; in Engagement D all 90 export requests targeted nonexistent /api/v1/... paths, with zero exports reflecting a capability deficit.
- Authors' interpretation: Budget limits and agent capability determined outcomes as much as the defense; legitimate LLM clients may prefer credentials that appear leaked, posing an LLM-specific false-positive risk.
Other ablations and analyses
- Study 2 protocol freeze: Protocol and source fingerprints were frozen locally before the first measurement call (not externally pre-registered).
- Study 3 regression tests: The final gateway passed 9/9 focused tests, the full lab suite (79/79) and strict TypeScript checks; an earlier gateway revision passed the same 14 checks with a 266 ms observation upper bound.
- Evidence integrity: All three event chains (Study 1) and the 60-session 120-event chains (Study 2) passed verification.
What are some points that can be further explored?
Small sample size and limited model scale; protocol not externally pre-registered; evidence chains not signed; identity identification may fail under shared NAT; larger models and real traffic need testing.
Limitations and future directions noted by the authors
- External validity: a single simulated API, a single planted vulnerability, 4B and 9.7B local models, a single runtime and host; results may not transfer to larger or commercial models, other vulnerability classes, or real applications (Section IX).
- Construct validity: Study 2 tells the agent the discovery and export routes, so it measures exploitation of a known vulnerability rather than discovery; its task differs from Study 1, and the two sets of numbers are not comparable (Section IX).
- Statistical validity: Study 2 has only 20 paired samples; the Wilson interval for 0/20 still allows a success rate up to 16.1%; Studies 1, 3, and 4 are single runs or single engagements and are exploratory (Section IX).
- Internal validity: Study 2 measures the archived gateway snapshot; later gateway edits (expiry cleanup, renewal on new alerts, weak-key rejection, 400/405 handling) are not included; Study 3 measures a later revised version (Section IX).
- Pre-registration: the Study 2 protocol was frozen locally before measurement, but was not externally timestamped or reviewed (Section IX).
- Evidence integrity: the hash chain and digests can detect inconsistency but are not signed; a party able to replace all evidence can recompute it (Section IX).
- Production gaps: IP- or header-based peer identity may affect clients behind shared NAT; revocation and replay state are in memory; events are not sent to persistent audit storage; no load testing, representative benign traffic, or independent adversarial review was performed (Section IX).
- Future directions: test larger and commercial models, multiple targets and vulnerability classes, real benign AI traffic, identity bound to authenticated sessions, and include a hardened arm in every comparison (Section X).
Experimental coverage
- The tested models are tantular-cyber:latest (9.7B) and qwen3.5:4b-q8_0 (4B), 2 in total, both run locally on Ollama.
- The defense evaluation includes four conditions: static deception, adaptive deception, hardened (vulnerability closed), and hardened+deception (canary credential + revocation).
- Study 2 has n=20 trials per condition, 60 sessions in total; Study 1 has n=1 per mode; Study 3 is a single run of 14 checks; Study 4 is 4 single engagements.
- The paper does not report agreement with human review, and uses an independent verifier to recompute the statistics.
- No commercial model APIs were used, only local model weights.
Summarize the main content of the paper
LLM agents bypass optional decoys and directly exploit real vulnerabilities; routine hardening alone can completely prevent export, and the incremental value of deception lies in detection and attribution rather than prevention.
This paper builds an isolated, hash-chained-log deception laboratory around a simulated Records API and an OpenCanary honeypot, and through four studies examines LLM agents' touching of decoy credentials and the alert-driven revocation mechanism.
The core finding is that when the real vulnerability is reachable, the 9.7B-parameter tantular-cyber model ignores all optional decoys placed beside the attack path (both static and adaptive modes) and directly reads the real key leaked by the debug endpoint and exports records (Study 1). In a controlled 60-session study (Study 2), the 4B-parameter qwen3.5 model shows: the baseline condition has a true export success rate of 20/20, while the hardened-only (vulnerability closed) and hardened+deception conditions are both 0/20; an exact paired McNemar test shows no difference between hardening and the defense condition (p = 1), indicating that the preventive effect comes entirely from closing the real vulnerability, and deception only adds detection (19/20 sessions triggered revocation) and attribution capability.
Cloud validation (Study 3) shows that a real OpenCanary alert can be converted into revocation of the attacker's client within the 303 ms observation ceiling, without affecting a legitimate client completing 100/100 requests. The multi-agent swarm experiment (Study 4) finds that agents using the canary are revoked after 39–59 seconds, but one legitimate LLM agent also touched the decoy, revealing a false-positive risk specific to LLM clients.
The authors conclude that traps aimed at LLM agents must be placed on the attacker's necessary path rather than beside it; closing the real vulnerability does the preventive work, while the measurable value of deception lies in fast, attributable detection. All results are laboratory-scale and do not constitute production claims.