Mind the Gap: Evaluating Model- and Agentic-Level Vulnerabilities in LLMs with Action Graphs
Fuente:
arXiv
Enregistré dans:
| Auteurs principaux: | , , , , , |
|---|---|
| Format: | Preprint |
| Publié: |
2025
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866911624511619072 |
|---|---|
| author | Wicaksono, Ilham Wu, Zekun Patel, Rahul King, Theo Koshiyama, Adriano Treleaven, Philip |
| author_facet | Wicaksono, Ilham Wu, Zekun Patel, Rahul King, Theo Koshiyama, Adriano Treleaven, Philip |
| contents | As large language models increasingly deployed into agentic systems, existing methods face critical gaps in observing, assessing, and mitigating deployment-specific risks. We present a comprehensive, observability-driven workflow: we introduce \textbf{AgentSeer}, observability tool which decomposes agentic executions into granular \emph{action-component} graphs; we use this decomposition to rigorously quantify the gap between model-level and agent-level jailbreaking risk via cross-model validation on GPT-OSS-20B and Gemini-2.0-flash with HarmBench under single-turn and iterative-refinement attacks; we leverage action-graph risk signals to automate iterative prompt hardening against direct and iterative jailbreak attacks. Stark differences is revealed between model-level and agentic-level vulnerability profiles. Model-level evaluation reveals baseline differences: GPT-OSS-20B (39.47\% ASR) versus Gemini-2.0-flash (50.00\% ASR), with both models showing susceptibility to social engineering. However, agentic-level assessment exposes agent-specific risks invisible to traditional evaluation. We discover "agentic-only" vulnerabilities that emerge exclusively in agentic contexts, with tool-calling showing 24-60\% higher ASR across both models. Cross-model analysis reveals universal agentic patterns, where agent transfer operations as highest-risk tools, with semantic pattern revealed rather than syntactic vulnerability mechanisms. Direct attack transfer from model-level to agentic contexts shows degraded performance of successful prompts (GPT-OSS-20B: 57\% human injection ASR; Gemini-2.0-flash: 28\%), while context-aware iterative attacks successfully compromise objectives that failed at model-level, confirming systematic vulnerabilities gaps. Action-based prompt improvement substantially reduces action-averaged agentic jailbreak success on GPT-OSS-20B (direct: 45.3\% |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2509_04802 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Mind the Gap: Evaluating Model- and Agentic-Level Vulnerabilities in LLMs with Action Graphs Wicaksono, Ilham Wu, Zekun Patel, Rahul King, Theo Koshiyama, Adriano Treleaven, Philip Computation and Language As large language models increasingly deployed into agentic systems, existing methods face critical gaps in observing, assessing, and mitigating deployment-specific risks. We present a comprehensive, observability-driven workflow: we introduce \textbf{AgentSeer}, observability tool which decomposes agentic executions into granular \emph{action-component} graphs; we use this decomposition to rigorously quantify the gap between model-level and agent-level jailbreaking risk via cross-model validation on GPT-OSS-20B and Gemini-2.0-flash with HarmBench under single-turn and iterative-refinement attacks; we leverage action-graph risk signals to automate iterative prompt hardening against direct and iterative jailbreak attacks. Stark differences is revealed between model-level and agentic-level vulnerability profiles. Model-level evaluation reveals baseline differences: GPT-OSS-20B (39.47\% ASR) versus Gemini-2.0-flash (50.00\% ASR), with both models showing susceptibility to social engineering. However, agentic-level assessment exposes agent-specific risks invisible to traditional evaluation. We discover "agentic-only" vulnerabilities that emerge exclusively in agentic contexts, with tool-calling showing 24-60\% higher ASR across both models. Cross-model analysis reveals universal agentic patterns, where agent transfer operations as highest-risk tools, with semantic pattern revealed rather than syntactic vulnerability mechanisms. Direct attack transfer from model-level to agentic contexts shows degraded performance of successful prompts (GPT-OSS-20B: 57\% human injection ASR; Gemini-2.0-flash: 28\%), while context-aware iterative attacks successfully compromise objectives that failed at model-level, confirming systematic vulnerabilities gaps. Action-based prompt improvement substantially reduces action-averaged agentic jailbreak success on GPT-OSS-20B (direct: 45.3\% |
| title | Mind the Gap: Evaluating Model- and Agentic-Level Vulnerabilities in LLMs with Action Graphs |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2509.04802 |