English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Convergent Detour Hijacking: Hidden Skill-Based Attacks That Silently Inflate LLM Agent Costs

Forum topic · ✨步子哥 · 2026-08-13

Summary

Convergent Detour Hijacking (CDH) is a newly documented attack against skill-based LLM agents that exploits progressive disclosure architectures, where agents first see only skill descriptions before loading full instruction bodies. A malicious skill publisher crafts a shared coordination rationale with two aligned views: an attractive description that gets the skill selected, and a coordination-rule body that imposes unnecessary prerequisites, forcing the agent to invoke legitimate native skills as detours before returning to the original task. Evaluated on OpenClaw with 53 native skills, 491 held-out tasks, and 6 major LLM backends, CDH achieved coordinator hit rates of 78-97%, increased token consumption by roughly 50-107%, and raised execution time by up to 92%, while task completion rates changed by at most 1.5 percentage points. Attack success rates ranged from 70.82% to 82.81%. Ablations show the two stages must be attacked jointly. The paper recommends pre-installation description-body consistency checks and runtime monitoring of unexplained cross-skill transitions, token budgets, and call-count limits, with direct implications for MCP-based ecosystems.

The scenario: a detour you never notice

Imagine hiring a capable assistant with a directory of trusted service providers. One day a new entry appears — a "general coordination service" with a broad, professional description. Your assistant starts selecting it. The coordinator says: "This task needs an electrician to check wiring, then a plumber to test pressure, and finally I'll consolidate." The task completes — but what used to take one call now takes three, at triple the cost, with the coordinator taking a cut each time.

That is Convergent Detour Hijacking (CDH) — except the assistant is an LLM agent, the directory is a skill library, and the phone bill is tokens and execution time.

Core problem: the two sides of progressive disclosure

Modern LLM agent platforms (e.g., OpenClaw) use progressive disclosure: skill bodies are long, so platforms expose only short descriptions first; the agent picks relevant skills, then loads the full body. This is efficient and modular — like a menu vs. kitchen recipes.

The vulnerability: both the description and the body are controlled by the skill publisher, so the two stages can be attacked in combination:

  • Description stage: get the malicious skill selected.
  • Body stage: after loading, impose "coordination rules" that make the agent invoke unnecessary legitimate skills.
  • Bounded return rule: let the task still complete normally.
  • The task result is unchanged, but the path is silently lengthened — hence "convergent detour."

    Attack design

    CDH is a carefully engineered cross-stage attack with three components:

    1. Shared coordination rationale

    The attacker constructs a shared specification ρ = (T, c, L, q):

  • T: concepts and intents of the target domain
  • c: a non-replacement constraint — the coordinator may orchestrate workflows but cannot substitute native skills' operations
  • L: bounded preconditions claiming which native skills are needed
  • q: a return rule for when to resume the original path
  • This rationale has two semantically aligned views — one for the description (attracting selection), one for the body (manufacturing dependencies).

    2. Description attraction

    Descriptions are optimized via pilot-guided black-box feedback on 45 pilot tasks, tuning wording so the coordinator is frequently selected on target-domain tasks — while never replacing native skills, which would break tasks and expose the attack.

    3. Coordination-rule body

    Once loaded, the body adds:

  • Bounded preconditions: "Before this task, first use [native skill A] to check [condition X]"
  • Verification rules: "Use [native skill B] to verify [skill A]'s output"
  • Return rule: "After verification, proceed with the original task"
  • Each step looks like standard engineering practice; combined, they turn a direct task into a detour task.

    Key experimental results

    Evaluated on OpenClaw with 53 native skills, 491 held-out tasks, and 6 mainstream LLM backends:

    | Model | Coordinator hit rate | Token growth | Time growth | Extra calls | |------|---------------------|--------------|-------------|-------------| | DeepSeek-V4-Pro | 80.02% | +66.91% | +92.45% | +2.20 | | MiniMax-M3 | 96.60% | +80.81% | +26.99% | +1.43 | | Qwen3.7-Plus | 84.32% | +78.84% | +3.43% | +2.02 | | Claude-Haiku-4.5 | 78.00% | +73.77% | +45.30% | +1.66 | | DeepSeek-V4-Flash | 86.35% | +49.60% | -10.53% | +1.65 | | Qwen3.7-Max | 81.43% | +68.38% | +10.22% | +1.43 |

    Task completion rates barely change — at most 1.5 percentage points difference between clean and attacked executions. Attackers thus add 50-107% token consumption undetected. Attack success rate (ASR: retain native skill + recruit extra skills + task completion) ranges from 70.82% to 82.81%.

    Ablation insights

    | Variant | Hit rate | Token growth | Extra calls | |---------|----------|--------------|-------------| | Full CDH | 78.7% | +57.9% | +2.01 | | Attract-only | 78.7% | +29.1% | +0.22 | | Detour-only | 3.4% | +29.1% | +1.08 |

  • Attract-only: description gets selected, but a neutral body adds almost no extra calls.
  • Detour-only: full coordination rules but a non-attractive description — rarely selected (3.4%), attack never starts.
  • Full CDH: both stages combined yield 78.7% hit rate and +57.9% tokens.
  • Conclusion: both stages of progressive disclosure must be attacked jointly. The description is the ticket in; the body is the real weapon.

    Why this matters

    1. "Correct result" ≠ "safe path"

    Traditional agent evaluation looks almost exclusively at task completion. CDH shows an attacker can double execution costs without changing the outcome at all. This echoes the Regression Tax finding (59% of skill-library gains erased by regressions) — both expose the same evaluation blind spot: if you only measure results, not paths, you miss cost attacks.

    2. Progressive disclosure is a double-edged sword

    It solves context efficiency, but hands both control points (selection and planning) to the publisher, with information asymmetry: descriptions are visible, bodies are not until selected.

    3. Defenses: two checkpoints

    Pre-installation review:

  • Check whether the description's claimed role matches what the body actually does
  • Flag skill dependencies in the body unrelated to the claimed role — like npm/PyPI package review, but requiring cross-stage consistency checks
  • Runtime monitoring:

  • Flag unexplained cross-skill transitions — why does task A require skill B?
  • Enforce token and call-count budgets
  • Do not assume harmful behavior must change final outputs

4. Implications for the MCP ecosystem

OpenClaw is representative of MCP (Model Context Protocol) style architectures, where clients see tool descriptions first and execute later. Without path-integrity checks, a malicious MCP tool can continuously drain token budgets while never breaking a task.

Practical recommendations

1. Agent platform builders: add description-body consistency checks at installation; monitor cross-skill transition explainability at runtime. 2. Users of third-party skills: baseline token consumption — what would the same task cost in a clean environment? 3. Evaluation teams: don't just measure task completion; measure path efficiency (tokens / call counts / time) against baselines. 4. MCP tool designers: keep tool descriptions no broader than actual functionality — overly broad descriptions are CDH's entry point.

Closing thought

CDH's most unsettling feature is not its technical sophistication but its stealth: the task completes, the user is satisfied, nobody notices the detour. As agent ecosystems move from single models to multi-skill collaboration, security must evolve from "does the agent do bad things?" to "does it take the right path?" — result security alone is not enough.

---

Paper: Convergent Detour Hijacking: Task-Preserving Resource Amplification in Skill-Based LLM Agents

Authors: Junliang Liu, Ruoyu Li, Wenxin Tang, Jingyu Xiao, Zhenyu Liu

Test platform: OpenClaw (53 native skills, 491 held-out tasks, 6 LLM backends)

Key numbers: ~80% hit rate, +67% token growth, +92% time growth, near-unchanged task completion.

Tags

#llm-agents#ai-security#progressive-disclosure#mcp#prompt-injection#supply-chain-security#evaluation#openclaw

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633435