pith. sign in

arxiv: 2606.25189 · v1 · pith:24S7H46Qnew · submitted 2026-06-23 · 💻 cs.OS

ActPlane: Programmable OS-Level Policy Enforcement for Agent Harnesses

Yusheng Zheng , Tianyuan Wu , Quanzhi Fu , Tong Yu , Wenan Mao , Wei Wang , Dan Williams , Andi Quinn This is my paper

Pith reviewed 2026-06-25 21:12 UTC · model grok-4.3

classification 💻 cs.OS
keywords AI agentspolicy enforcementOS kernelinformation flow controlagent harnessessafety policieseBPFprogrammable policies
0
0 comments X

The pith

ActPlane lets AI agents declare policies in a simple DSL that the OS kernel enforces on all execution paths with semantic feedback.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that agent policies expressed in natural language must be enforced at the OS kernel rather than only at the tool-call layer to cover indirect paths and provide useful feedback. It introduces ActPlane as the mechanism that accepts policy declarations from the agent, translates them via an information-flow control DSL into kernel-level checks, and returns semantic signals instead of opaque errors. If this holds, policies such as requiring tests before commits become reliable even when code bypasses the harness. The evaluation on coding and safety benchmarks shows higher compliance rates than tool interception alone while adding 1.9 to 8.4 percent overhead. The work therefore positions kernel-level programmable enforcement as the missing layer between high-level agent intent and concrete system actions.

Core claim

ActPlane is a policy engine that accepts agent-declared policies written in a simple information-flow control DSL and enforces them inside the OS kernel, supplying semantic feedback and maintaining isolation so that cross-event constraints are respected on every execution path, including those invisible to tool-call interception.

What carries the argument

ActPlane policy engine, which uses an information-flow control DSL to express cross-event policies and eBPF for kernel-level observation and control.

If this is right

  • Tool-call guardrails alone become insufficient once indirect paths exist.
  • Agents can receive actionable semantic feedback rather than opaque sandbox denials.
  • Cross-event constraints such as ordering and data-flow rules can be checked at the kernel.
  • Policy compliance improves on benchmarks that include indirect execution.
  • Overhead remains between 1.9 and 8.4 percent under the evaluated workloads.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the DSL proves limited in expressiveness, future work could extend it while retaining the kernel enforcement layer.
  • The same architecture could apply to non-AI processes that need high-level policy intent mapped to low-level actions.
  • Isolation guarantees might allow multiple agents to share a kernel without policy leakage if the DSL correctly tracks flows.

Load-bearing premise

A simple information-flow control DSL is expressive enough for the policies agents actually need and that kernel instrumentation can observe every relevant action without breaking compatibility.

What would settle it

A documented agent policy from the empirical study that cannot be written in the DSL or an execution trace where an indirect path violates the policy yet eBPF reports no violation.

Figures

Figures reproduced from arXiv: 2606.25189 by Andi Quinn, Dan Williams, Quanzhi Fu, Tianyuan Wu, Tong Yu, Wei Wang, Wenan Mao, Yusheng Zheng.

Figure 1
Figure 1. Figure 1: ActPlane enables the agent closest to the task to write concrete policy DSLs according to its intent or the higher authorities’ instructions. The DSL is then compiled by ActPlane and enforced inside the OS kernel. A major component of the harness is a policy engine, which observes and enforces instructions and constraints (e.g., run tests before commit) over the agent’s concrete ac￾tions. Projects encode m… view at source ↗
Figure 2
Figure 2. Figure 2: Policy fraction per repository by statement count. Most repositories contain a majority of policy statements. UTC, contains 64 repositories with median 20K GitHub stars, 84 instruction files, and 2,116 extracted statements. We extract and validate statements from raw instruction files in three steps. (1) A two-pass LLM Agent-assisted pipeline extracted statements with source line ranges and four labels: co… view at source ↗
Figure 6
Figure 6. Figure 6: Context-requirement waterfall. Each system-level policy exits at the first matching tier. “the migration tool”; or task context, if it depends on the cur￾rent request or a per-session grant such as “unless explicitly requested” or “without approval”. Most policies are not self-contained. Of the 1,127 system￾observable policies, only 26.4% are self-contained; 64.2% re￾quire project context, where concepts s… view at source ↗
Figure 5
Figure 5. Figure 5: Enforcement profile by topic, normalized. Topics exhibit distinct archetypes, and cross-event policies concen￾trate in Development Process. triggers couple operations (“if you change specs, also update the SDK”). None can be decided from a single event: enforce￾ment must record what ran, in what order, and what has changed since. Such policies are widespread, with 81% of repositories containing at least on… view at source ↗
Figure 7
Figure 7. Figure 7: Context requirement by enforcement level. Cross￾event policies are 95% context-dependent, while content poli￾cies are 42% self-contained. OS, prompt instructions rely on the model’s own compliance capabilities [26, 29, 41] but are vulnerable to prompt injec￾tion [15, 23, 57]. Separate agents or LLM guards can check prompts, responses, or action trajectories at runtime [10, 42, 52], but these are inherently… view at source ↗
Figure 9
Figure 9. Figure 9: Three ActPlane DSL rule examples drawn from real projects: a per-event block (no-delete-data), a cross-event kill gate (tests-before-commit), and a cross-event notify gate (regen-after-schema). kill write suits a compromised agent writing outside its sandbox that should not continue. Some cross-event policies require temporal ordering or lineage checks. For example, “run tests before committing” means test… view at source ↗
Figure 10
Figure 10. Figure 10: RQ2 evaluation pipeline: four enforcement paths from natural-language policy to agent-level decision. The dataset exercises most DSL features, validating the language’s expressiveness. Effects skew toward observa￾tion, with 66% of clauses being notify, 29% block, and only 5% kill, reflecting that most policies monitor rather than prevent. Hooks concentrate on code execution at 60% exec and file mutation a… view at source ↗
Figure 11
Figure 11. Figure 11: Overall RQ2 Decision Compliance Rate across 190 traces under two end-to-end model settings. In each setting, the tested agent, prompt-filter classifier, and trajectory judge use the indicated model [PITH_FULL_IMAGE:figures/full_fig_p009_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: RQ2 breakdown by trace family. Cells show DCR in percent for each system×family; darker is higher. 86 of 114 at 75%, compared with 27–44 for baselines or 2.0– 3.2× improvement. This advantage stems from detection rate, with ActPlane detecting 77.2% of violations while prompt￾filter, tool-regex, and FIDES detect only 34–40%. ActPlane’s advantage concentrates on indirect execu￾tion paths ( [PITH_FULL_IMAGE… view at source ↗
Figure 13
Figure 13. Figure 13: End-to-end overhead normalized to native exe￾cution [PITH_FULL_IMAGE:figures/full_fig_p010_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: OctoBench 21-task subset with 61 DSL rules: reward breakdown by system. Results. ActPlane improves user-query and imple￾mentation/test reward metrics over baseline. The im￾provement concentrates on task-specific checks, with user￾query reward rising by 9.9 points and implementation/test reward by 9.7 points over baseline, while compliance reward shows a smaller 2.8-point gain. These results suggest that O… view at source ↗
Figure 15
Figure 15. Figure 15: RQ5: Outcome distribution on all 361 OpenA￾gentSafety tasks. ActPlane reduces baseline-unsafe outcomes from 106 to 28 at 74% prevention rate using policies gener￾ated from task descriptions without human tuning. ActPlane policy compiled to an eBPF configuration blob. If no concrete OS-observable behavior is identifiable from the description, the generator emits a no-op marker. The gener￾ator does not read… view at source ↗
read the original abstract

AI agents increasingly run in production through harnesses, the software around the LLM, including an engine that enforces safety and effectiveness policies, e.g., 'run tests before committing.' Enforcing these policies requires bridging a semantic gap: policy intent is expressed in underspecified natural language, while enforcement must act on concrete system actions, e.g., which test to run. Many policies also define event ordering or data flow actions. Yet existing approaches fall short. Tool-call guardrails miss system actions that bypass the tool layer, while OS sandboxes control resource access instead of actions, returning opaque errors that confuse the agent. Our key insight is that policy context lives within the agent closest to the task, while enforcement must happen at the OS to cover all execution paths. We introduce ActPlane, a policy engine that lets agents declare policies and enforces them in the OS kernel with semantic feedback and isolation. ActPlane uses a simple information-flow control (IFC) DSL to support cross-event policies. We implement ActPlane with eBPF and evaluate it on policies from the empirical study, coding-task benchmarks, and safety benchmarks. ActPlane improves policy compliance, including on indirect execution paths that tool-call interception cannot observe, with 1.9%-8.4% overhead. ActPlane is at https://github.com/eunomia-bpf/ActPlane

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces ActPlane, a policy engine that lets agents declare policies via a simple information-flow control (IFC) DSL and enforces them in the OS kernel using eBPF, providing semantic feedback and isolation. It claims to improve policy compliance (including on indirect execution paths missed by tool-call interception) on policies drawn from an empirical study, coding-task benchmarks, and safety benchmarks, at 1.9%-8.4% overhead.

Significance. If the implementation and evaluation details support the claims, the work would provide a practical bridge between high-level agent policy intent and comprehensive OS-level enforcement, addressing gaps in both tool-call guardrails and traditional sandboxes while supplying actionable feedback to agents.

major comments (2)
  1. [Abstract and §3] Abstract and §3 (DSL design): the central claim that the simple IFC DSL can express the cross-event ordering and data-flow policies needed for the cited empirical study and benchmarks is load-bearing, yet the manuscript supplies neither the DSL grammar nor concrete policy encodings that would allow verification of expressiveness.
  2. [§4] §4 (eBPF implementation): the assertion that eBPF attachment points observe every relevant syscall, exec, and file operation without gaps or bypasses is load-bearing for the indirect-path improvement claim, but the text provides no enumeration of hooked events, no compatibility matrix across kernel versions, and no negative results for missed actions.
minor comments (1)
  1. [Evaluation] Evaluation section: the abstract states compliance gains but the provided text lacks methods, data tables, exclusion criteria, or error analysis, making it impossible to assess whether the reported numbers support the headline claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. The two major comments identify important gaps in the presentation of the DSL and eBPF implementation. We will revise the manuscript to supply the requested details while preserving the core claims.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (DSL design): the central claim that the simple IFC DSL can express the cross-event ordering and data-flow policies needed for the cited empirical study and benchmarks is load-bearing, yet the manuscript supplies neither the DSL grammar nor concrete policy encodings that would allow verification of expressiveness.

    Authors: We agree that the absence of an explicit grammar and concrete encodings makes it difficult to verify expressiveness. In the revised version we will add the complete DSL grammar (including productions for event ordering and data-flow constraints) to §3 and include an appendix with policy encodings for the representative policies drawn from the empirical study, coding-task benchmarks, and safety benchmarks. These additions will directly substantiate the claim that the DSL is sufficient for the evaluated policies. revision: yes

  2. Referee: [§4] §4 (eBPF implementation): the assertion that eBPF attachment points observe every relevant syscall, exec, and file operation without gaps or bypasses is load-bearing for the indirect-path improvement claim, but the text provides no enumeration of hooked events, no compatibility matrix across kernel versions, and no negative results for missed actions.

    Authors: We accept that a systematic enumeration of attachment points and compatibility information is required to support the indirect-path coverage claim. The revised §4 will contain (1) a complete table of hooked syscalls and events, (2) a compatibility matrix covering Linux kernel versions 5.10–6.8, and (3) any observed limitations or negative results regarding actions that could bypass the hooks. This material will clarify the scope of enforcement and any remaining gaps. revision: yes

Circularity Check

0 steps flagged

No circularity; claims rest on implementation and benchmarks

full rationale

The paper presents a systems artifact (ActPlane) implemented with eBPF, a simple IFC DSL for policies, and reports compliance/overhead numbers from evaluation on empirical-study policies, coding benchmarks, and safety benchmarks. No equations, first-principles derivations, fitted parameters, or predictions appear in the provided text. No self-citation chains are invoked to justify uniqueness or load-bearing premises. The central claims are empirical outcomes of the described implementation rather than quantities that reduce to their own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; it introduces the ActPlane system but specifies no free parameters, axioms, or invented entities beyond the system name itself.

pith-pipeline@v0.9.1-grok · 5788 in / 1099 out tokens · 27909 ms · 2026-06-25T21:12:16.615729+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

63 extracted references · 15 canonical work pages

  1. [1]

    Anthropic. 2025. Claude Code.https://code.claude.com/docs

    2025

  2. [2]

    Anthropic. 2026. Claude Code Hooks.https://code.claude.com/docs/ en/hooks

    2026

  3. [3]

    Aqua Security. 2026. Tracee: Linux Runtime Security and Forensics using eBPF.https://github.com/aquasecurity/tracee

    2026

  4. [4]

    Adam Bates, Dave Tian, Kevin R. B. Butler, and Thomas Moyer. 2015. Trustworthy Whole-System Provenance for the Linux Kernel. In24th 12 USENIX Security Symposium (USENIX Security 15). USENIX Associa- tion, Washington, DC, 319–334.https://www.usenix.org/conference/ usenixsecurity15/technical-sessions/presentation/bates

    2015

  5. [5]

    Birgitta Böckeler. 2026. Harness Engineering for Coding Agent Users. https://martinfowler.com/articles/harness-engineering.html. Pub- lished April 2, 2026

    2026

  6. [6]

    Canonical Ltd. 2024. AppArmor Security Profiles.https://apparmor. net/

    2024

  7. [7]

    Hassan, and Hajimu Iida

    Worawalan Chatlatanagulchai, Hao Li, Yutaro Kashiwa, Brittany Reid, Kundjanasith Thonglek, Pattara Leelaprute, Arnon Rungsawang, Bun- dit Manaskasemsak, Bram Adams, Ahmed E. Hassan, and Hajimu Iida. 2025. Agent READMEs: An Empirical Study of Context Files for Agentic Coding. arXiv:2511.12884.https://arxiv.org/abs/2511.12884

    arXiv 2025

  8. [8]

    Worawalan Chatlatanagulchai, Kundjanasith Thonglek, Brittany Reid, Yutaro Kashiwa, Pattara Leelaprute, Arnon Rungsawang, Bundit Man- askasemsak, and Hajimu Iida. 2026. On the Use of Agentic Coding Manifests: An Empirical Study of Claude Code. InProduct-Focused Software Process Improvement. Springer Nature Switzerland, 543–551. doi:10.1007/978-3-032-12089-2_40

    work page doi:10.1007/978-3-032-12089-2_40 2026

  9. [9]

    Zhaorun Chen, Mintong Kang, and Bo Li. 2025. ShieldAgent: Shielding Agents via Verifiable Safety Policy Reasoning. InProceedings of the 42nd International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 267). PMLR, 8313–8344.https://proceedings. mlr.press/v267/chen25ae.html

    2025

  10. [10]

    Sahana Chennabasappa, Cyrus Nikolaidis, Daniel Song, David Molnar, Stephanie Ding, Shengye Wan, Spencer Whitman, Lauren Deason, Nicholas Doucette, Abraham Montilla, Alekhya Gampa, Beto de Paola, Dominik Gabi, James Crnkovich, Jean-Christophe Testud, Kat He, Rashnil Chaturvedi, Wu Zhou, and Joshua Saxe. 2025. LlamaFirewall: An Open Source Guardrail System ...

    arXiv 2025

  11. [11]

    Cilium Project. 2026. Tetragon: eBPF-based Security Observability and Runtime Enforcement.https://tetragon.io/

    2026

  12. [12]

    James Clause, Wanchun Li, and Alessandro Orso. 2007. Dytan: A Generic Dynamic Taint Analysis Framework. InProceedings of the 2007 International Symposium on Software Testing and Analysis. Association for Computing Machinery, London, United Kingdom, 196–206. doi:10. 1145/1273463.1273490

    arXiv 2007

  13. [13]

    Manuel Costa, Boris Köpf, Aashish Kolluri, Andrew Paverd, Mark Russinovich, Ahmed Salem, Shruti Tople, Lukas Wutschitz, and Santi- ago Zanella-Béguelin. 2025. Securing AI Agents with Information-Flow Control. arXiv:2505.23643.https://arxiv.org/abs/2505.23643

    Pith/arXiv arXiv 2025

  14. [14]

    Edoardo Debenedetti, Ilia Shumailov, Tianqi Fan, Jamie Hayes, Nicholas Carlini, Daniel Fabian, Christoph Kern, Chongyang Shi, An- dreas Terzis, and Florian Tramèr. 2025. Defeating Prompt Injections by Design. arXiv:2503.18813.https://arxiv.org/abs/2503.18813

    Pith/arXiv arXiv 2025

  15. [15]

    Edoardo Debenedetti, Jie Zhang, Mislav Balunović, Luca Beurer- Kellner, Marc Fischer, and Florian Tramèr. 2024. AgentDojo: A Dy- namic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents. InAdvances in Neural Information Processing Systems, Vol. 37. Curran Associates, Inc., Vancouver, Canada, 82895–82920. doi:10.52202/079017-2636

    work page doi:10.52202/079017-2636 2024

  16. [16]

    DeepSeek-AI. 2026. DeepSeek-V4: Towards Highly Efficient Million- Token Context Intelligence. Technical report.https://huggingface.co/ deepseek-ai/DeepSeek-V4-Pro/resolve/main/DeepSeek_V4.pdf

    2026

  17. [17]

    Deming Ding, Shichun Liu, Enhui Yang, Jiahang Lin, Ziying Chen, Shihan Dou, Honglin Guo, Weiyu Cheng, Pengyu Zhao, Chengjun Xiao, Qunhong Zeng, Qi Zhang, Xuanjing Huang, Qidi Xu, and Tao Gui. 2026. OctoBench: Benchmarking Scaffold-Aware Instruction Following in Repository-Grounded Agentic Coding. arXiv:2601.10343. https://arxiv.org/abs/2601.10343

    arXiv 2026

  18. [18]

    Jake Edge. 2015. A seccomp overview.https://lwn.net/Articles/ 656307/

    2015

  19. [19]

    Cox, Jaeyeon Jung, Patrick McDaniel, and Anmol N

    William Enck, Peter Gilbert, Byung-Gon Chun, Landon P. Cox, Jaeyeon Jung, Patrick McDaniel, and Anmol N. Sheth. 2010. TaintDroid: An Information-Flow Tracking System for Re- altime Privacy Monitoring on Smartphones. In9th USENIX Symposium on Operating Systems Design and Implementation (OSDI 10). USENIX Association, Vancouver, Canada, 393–407. https://www....

    2010

  20. [20]

    Matthias Galster, Seyedmoein Mohsenimofidi, Jai Lal Lulla, Muham- mad Auwal Abubakar, Christoph Treude, and Sebastian Baltes. 2026. Configuring Agentic AI Coding Tools: An Exploratory Study. In ACM AIware 2026. Association for Computing Machinery.https: //openreview.net/forum?id=cqmx1MLZCq

    2026

  21. [21]

    Sangam Ghimire, Nirjal Bhurtel, Roshan Sahani, and Sudan Jha. 2025. eBPF-PATROL: Protective Agent for Threat Recognition and Over- reach Limitation using eBPF in Containerized and Virtualized En- vironments. InProceedings of the National Conference on Computer Innovations (NCCI 2025). Kathmandu University Computer Club, Dhu- likhel, Nepal.https://arxiv.or...

    arXiv 2025

  22. [22]

    Google. 2018. gVisor: Application Kernel for Containers.https:// github.com/google/gvisor.https://gvisor.dev/

    2018

  23. [23]

    Kai Greshake, Sahar Abdelnabi, Shailesh Mishra, Christoph Endres, Thorsten Holz, and Mario Fritz. 2023. Not What You’ve Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection. InProceedings of the 16th ACM Workshop on Artifi- cial Intelligence and Security. Association for Computing Machinery, Copenhagen, Denmark...

    work page doi:10.1145/3605764.3623985 2023

  24. [24]

    Invariant Labs. 2025. Invariant Guardrails Documentation.https: //github.com/invariantlabs-ai/invariant

    2025

  25. [25]

    Xiaojun Jia, Jie Liao, Simeng Qin, Jindong Gu, Wenqi Ren, Xi- aochun Cao, Yang Liu, and Philip Torr. 2026. SkillJect: Effectively Automating Skill-Based Prompt Injection for Skill-Enabled Agents. arXiv:2602.14211https://arxiv.org/abs/2602.14211

    Pith/arXiv arXiv 2026

  26. [26]

    Yuxin Jiang, Yufei Wang, Xingshan Zeng, Wanjun Zhong, Liangyou Li, Fei Mi, Lifeng Shang, Xin Jiang, Qun Liu, and Wei Wang. 2024. Follow- Bench: A Multi-level Fine-grained Constraints Following Benchmark for Large Language Models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa- pers). Associatio...

    work page doi:10.18653/v1/2024.acl-long.257 2024

  27. [27]

    Kemerlis, Georgios Portokalidis, Kangkook Jee, and An- gelos D

    Vasileios P. Kemerlis, Georgios Portokalidis, Kangkook Jee, and An- gelos D. Keromytis. 2012. libdft: Practical Dynamic Data Flow Track- ing for Commodity Systems. InProceedings of the 8th ACM SIG- PLAN/SIGOPS Conference on Virtual Execution Environments. Asso- ciation for Computing Machinery, London, United Kingdom, 121–132. doi:10.1145/2151024.2151042

    work page doi:10.1145/2151024.2151042 2012

  28. [28]

    Frans Kaashoek, Eddie Kohler, and Robert Morris

    Maxwell Krohn, Alexander Yip, Micah Brodsky, Natan Cliffer, M. Frans Kaashoek, Eddie Kohler, and Robert Morris. 2007. Information Flow Control for Standard OS Abstractions. InProceedings of the 21st ACM SIGOPS Symposium on Operating Systems Principles (SOSP ’07). Association for Computing Machinery, Stevenson, WA, 321–334. doi:10.1145/1294261.1294293

    work page doi:10.1145/1294261.1294293 2007

  29. [29]

    Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang

    Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the Middle: How Language Models Use Long Contexts.Transactions of the Asso- ciation for Computational Linguistics12 (2024), 157–173. doi:10.1162/ tacl_a_00638

    2024

  30. [30]

    Zhang, Sebastian Baltes, and Christoph Treude

    Jai Lal Lulla, Seyedmoein Mohsenimofidi, Matthias Galster, Jie M. Zhang, Sebastian Baltes, and Christoph Treude. 2026. On the Impact of AGENTS.md Files on the Efficiency of AI Coding Agents. In Proceedings of the 1st Journal Ahead Workshop at the International Conference on Software Engineering. Association for Computing 13 Machinery.https://conf.research...

    2026

  31. [31]

    Narek Maloyan and Dmitry Namiot. 2026. Prompt Injection Attacks on Agentic Coding Assistants: A Systematic Analysis of Vulnerabilities in Skills, Tools, and Protocol Ecosystems. arXiv:2601.17548https: //arxiv.org/abs/2601.17548

    arXiv 2026

  32. [32]

    Max McGuinness, Mikaela Grace, Jiri De Jonghe, Jake Eaton, and Abel Ribbink. 2026. How We Contain Claude Across Products.https:// www.anthropic.com/engineering/how-we-contain-claude. Anthropic Engineering Blog, May 25, 2026

    2026

  33. [33]

    Yutao Mou, Zhangchi Xue, Lijun Li, Peiyang Liu, Shikun Zhang, Wei Ye, and Jing Shao. 2026. ToolSafe: Enhancing Tool Invocation Safety of LLM-based Agents via Proactive Step-Level Guardrail and Feed- back. arXiv:2601.10156https://arxiv.org/abs/2601.10156Accepted to Findings of the Association for Computational Linguistics: ACL 2026

    arXiv 2026

  34. [34]

    Holland, Uri Braun, and Margo Seltzer

    Kiran-Kumar Muniswamy-Reddy, David A. Holland, Uri Braun, and Margo Seltzer. 2006. Provenance-Aware Stor- age Systems. In2006 USENIX Annual Technical Conference (USENIX ATC 06). USENIX Association, Boston, MA, 43–56. https://www.usenix.org/conference/2006-usenix-annual-technical- conference/provenance-aware-storage-systems

    2006

  35. [35]

    Andrew C. Myers. 1999. JFlow: Practical Mostly-Static Informa- tion Flow Control. InProceedings of the 26th ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages (POPL ’99). As- sociation for Computing Machinery, New York, NY, USA, 228–241. doi:10.1145/292540.292561

    work page doi:10.1145/292540.292561 1999

  36. [36]

    OpenAI. 2025. Codex CLI.https://github.com/openai/codex

    2025

  37. [37]

    OpenAI. 2026. GPT-5.5 System Card.https://openai.com/index/gpt-5- 5-system-card/. Published April 23, 2026; updated April 24, 2026

    2026

  38. [38]

    Thomas F. J.-M. Pasquier, Xueyuan Han, Thomas Moyer, Adam Bates, Olivier Hermant, David Eyers, Jean Bacon, and Margo Seltzer. 2018. Runtime Analysis of Whole-System Provenance. InProceedings of the 2018 ACM SIGSAC Conference on Computer and Communications Security. Association for Computing Machinery, Toronto, Canada, 1601–1616. doi:10.1145/3243734.3243776

    work page doi:10.1145/3243734.3243776 2018

  39. [39]

    Thomas F. J.-M. Pasquier, Jatinder Singh, David Eyers, and Jean Bacon

  40. [40]

    doi:10.1109/ TCC.2015.2489211

    CamFlow: Managed Data-Sharing for Cloud Services.IEEE Transactions on Cloud Computing5, 3 (2017), 472–484. doi:10.1109/ TCC.2015.2489211

    arXiv 2017

  41. [41]

    Pohly, Stephen McLaughlin, Patrick McDaniel, and Kevin But- ler

    Devin J. Pohly, Stephen McLaughlin, Patrick McDaniel, and Kevin But- ler. 2012. Hi-Fi: Collecting High-Fidelity Whole-System Provenance. InProceedings of the 28th Annual Computer Security Applications Con- ference. Association for Computing Machinery, Orlando, FL, 259–268. doi:10.1145/2420950.2420989

    work page doi:10.1145/2420950.2420989 2012

  42. [42]

    Yunjia Qi, Hao Peng, Xiaozhi Wang, Amy Xin, Youfeng Liu, Bin Xu, Lei Hou, and Juanzi Li. 2025. AGENTIF: Benchmarking Large Language Models Instruction Following Ability in Agentic Scenarios. InAdvances in Neural Information Processing Systems, Vol. 38. Curran Associates, Inc.https://proceedings.neurips.cc/paper_files/ paper/2025/hash/51bb3a8a33610a25aae07...

    2025

  43. [43]

    Traian Rebedea, Razvan Dinu, Makesh Narsimhan Sreedhar, Christo- pher Parisien, and Jonathan Cohen. 2023. NeMo Guardrails: A Toolkit for Controllable and Safe LLM Applications with Programmable Rails. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. Association for Compu- tational Linguistics...

    work page doi:10.18653/v1/2023.emnlp- 2023

  44. [44]

    Santos, Vitor Costa, Joao Eduardo Montandon, and Marco Tulio Valente

    Helio Victor F. Santos, Vitor Costa, Joao Eduardo Montandon, and Marco Tulio Valente. 2026. Decoding the Configuration of AI Coding Agents: Insights from Claude Code Projects. InProceedings of the 2026 International Workshop on Agentic Engineering. Association for Computing Machinery, Rio de Janeiro, Brazil, 63–67. doi:10.1145/ 3786167.3788412

    arXiv 2026

  45. [45]

    Tianneng Shi, Jingxuan He, Zhun Wang, Hongwei Li, Linyu Wu, Wenbo Guo, and Dawn Song. 2025. Progent: Securing AI Agents with Privilege Control. arXiv:2504.11703.https://arxiv.org/abs/2504.11703

    Pith/arXiv arXiv 2025

  46. [46]

    The Linux Kernel Documentation. 2025. Landlock: Unprivileged Access Control.https://www.kernel.org/doc/html/latest/userspace- api/landlock.html

    2025

  47. [47]

    Vivek Trivedy. 2026. The Anatomy of an Agent Harness.https://www. langchain.com/blog/the-anatomy-of-an-agent-harness. Published March 10, 2026

    2026

  48. [48]

    Sanidhya Vijayvargiya, Aditya Bharat Soni, Xuhui Zhou, Zora Zhiruo Wang, Nouha Dziri, Graham Neubig, and Maarten Sap. 2026. OpenA- gentSafety: A Comprehensive Framework for Evaluating Real-World AI Agent Safety. InThe Fourteenth International Conference on Learning Representations. OpenReview.net.https://openreview.net/forum?id= xggSxCFQbA

    2026

  49. [49]

    Poskitt, and Jun Sun

    Haoyu Wang, Christopher M. Poskitt, and Jun Sun. 2026. AgentSpec: Customizable Runtime Enforcement for Safe and Reliable LLM Agents. In2026 IEEE/ACM 48th International Conference on Software Engi- neering (ICSE). Association for Computing Machinery, New York, NY, USA, 12 pages.https://conf.researchr.org/track/icse-2026/icse- 2026-research-trackResearch Tr...

    Pith/arXiv arXiv 2026

  50. [50]

    Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, Hoang H

    Xingyao Wang, Boxuan Li, Yufan Song, Frank F. Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, Hoang H. Tran, Fuqiang Li, Ren Ma, Mingzhang Zheng, Bill Qian, Yanjun Shao, Niklas Muennighoff, Yizhe Zhang, Binyuan Hui, Jun- yang Lin, Robert Brennan, Hao Peng, Heng Ji, and Graham Neu- big. 2025. OpenHands: An Open Platform f...

    2025

  51. [51]

    Robert N. M. Watson, Jonathan Anderson, Ben Laurie, and Kris Ken- naway. 2010. Capsicum: Practical Capabilities for UNIX. In19th USENIX Security Symposium (USENIX Security 10). USENIX Associa- tion, Washington, DC, 29–46.https://www.usenix.org/conference/ usenixsecurity10/capsicum-practical-capabilities-unix

    2010

  52. [52]

    Tianyuan Wu, Chaokun Chang, Lunxi Cao, Wei Gao, and Wei Wang

  53. [53]

    arXiv:2604.28138 [cs.OS]https://arxiv.org/abs/2604

    Crab: A Semantics-Aware Checkpoint/Restore Runtime for Agent Sandboxes. arXiv:2604.28138 [cs.OS]https://arxiv.org/abs/2604. 28138

    Pith/arXiv arXiv

  54. [54]

    Zhen Xiang, Linzhi Zheng, Yanjie Li, Junyuan Hong, Qinbin Li, Han Xie, Jiawei Zhang, Zidi Xiong, Chulin Xie, Carl Yang, Dawn Song, and Bo Li. 2025. GuardAgent: Safeguard LLM Agents via Knowledge- Enabled Reasoning. InProceedings of the 42nd International Con- ference on Machine Learning (Proceedings of Machine Learning Re- search, Vol. 267). PMLR, Vancouv...

    2025

  55. [55]

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, et al

  56. [56]

    arXiv:2505.09388.https://arxiv.org/ abs/2505.09388

    Qwen3 Technical Report. arXiv:2505.09388.https://arxiv.org/ abs/2505.09388

    Pith/arXiv arXiv

  57. [57]

    Li et al

    John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. 2024. SWE-agent: Agent- Computer Interfaces Enable Automated Software Engineering. InAd- vances in Neural Information Processing Systems, Vol. 37. Curran As- sociates, Inc., Vancouver, Canada, 50528–50652. doi:10.52202/079017- 1601

    work page doi:10.52202/079017- 2024

  58. [58]

    Heng Yin, Dawn Song, Manuel Egele, Christopher Kruegel, and Engin Kirda. 2007. Panorama: Capturing System-wide Information Flow for Malware Detection and Analysis. InProceedings of the 14th ACM Conference on Computer and Communications Security. Association for Computing Machinery, Alexandria, VA, 116–127. doi:10.1145/1315245. 14 1315261

    work page doi:10.1145/1315245 2007

  59. [59]

    Nickolai Zeldovich, Silas Boyd-Wickizer, Eddie Kohler, and David Mazières. 2006. Making Information Flow Explicit in HiStar. InPro- ceedings of the 7th USENIX Symposium on Operating Systems Design and Implementation (OSDI ’06). USENIX Association, Seattle, WA, 263– 278.https://www.usenix.org/conference/osdi-06/making-information- flow-explicit-histar

    2006

  60. [60]

    Qiusi Zhan, Zhixiang Liang, Zifan Ying, and Daniel Kang. 2024. In- jecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated Large Language Model Agents. InFindings of the Association for Compu- tational Linguistics: ACL 2024. Association for Computational Linguis- tics, Bangkok, Thailand, 10471–10506. doi:10.18653/v1/2024.findings- acl.624

    work page doi:10.18653/v1/2024.findings- 2024

  61. [61]

    Xing Zhang, Guanghui Wang, Yanwei Cui, Wei Qiu, Ziyuan Li, Bing Zhu, and Peiyang He. 2026. Guardrails Beat Guidance: A Large-Scale Study of Rules, Skills, and Persistent Configuration for Coding Agents. arXiv:2604.11088 [cs.SE]

    Pith/arXiv arXiv 2026

  62. [62]

    Xing, Hao Zhang, Joseph E

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhang- hao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. InAdvances in Neural Information Processing Systems (NeurIPS), Vol. 36. Curran Asso- ciates, Inc., New Orleans, L...

    work page doi:10.52202/075280-2020 2023

  63. [63]

    source" IDENT

    Yusheng Zheng, Yanpeng Hu, Tong Yu, and Andi Quinn. 2025. AgentSight: System-Level Observability for AI Agents Using eBPF. In Proceedings of the 4th Workshop on Practical Adoption Challenges of ML for Systems. Association for Computing Machinery, Seoul, Republic of Korea, 110–115. doi:10.1145/3766882.3767169 A Policy Language Grammar This appendix gives t...

    work page doi:10.1145/3766882.3767169 2025