Executive reading · ~60 seconds
ToolPulse indicates that the agent’s logical contract does not complete isolation when local tools share observable physical resources.
Running the model remotely and the tool inside a sandbox reduces important surfaces, but it does not complete the threat model. The ToolPulse preprint describes a side channel in which an unprivileged local application observes microarchitectural traces left by shared tool execution and infers which workflow an agent is performing.
Status: Evidence Brief approved for publication. Editorial series: Papers we read for you. The text describes conditions, impact, and defensive controls without reproducing offensive procedures, vulnerable addresses, or exploitation instructions.
What the study measured
The authors evaluated 11 models in Codex and Claude Code workflows across three hardware platforms. In the controlled experiment, the tool-workflow classifier achieved F1 above 93% in every reported configuration. In two broader classification tasks, the reported accuracies were 87.77% and 91.39%.
The work also simulated noise on a shared server equivalent to approximately 20 users. According to the authors, inference remained viable in that scenario. These numbers belong to the paper’s experimental design; they are not a universal exploitation rate for any cloud, agent, operating system, or configuration.
The boundary of the finding matters as much as the result
ToolPulse does not describe a remote observer discovering prompts through the API. The threat hypothesis requires co-location on the same physical processor and shared executable pages or tool libraries. The observer is an unprivileged local process, but it must be in the right environment to capture the signal. Web-only agents and environments that do not share these conditions are outside the evaluated scope.
The result also does not prove access to the full prompt, file, or model response. What is inferred is the tool-flow pattern and, for specific tasks, an activity class. Turning that signal into impact depends on the context, workflow sensitivity, and the environment’s other controls.
Why the agent sandbox does not end the discussion
A sandbox governs capabilities such as files, network, processes, and allowed calls. The evaluated channel operates at another layer: shared physical resources and executable code. Therefore, a correct tool policy can coexist with a side-channel leak outside the agent’s logical contract.
The engineering principle is simple: isolation must cover the complete execution path. For sensitive workloads, that includes process, tenant, virtual-machine, host, and hardware boundaries — not only the permission declared to the model. [R2-C3]
Defensive controls proportional to risk
- Classify workflows by sensitivity before deciding whether they may share a host or CPU.
- For high-impact operations, prefer strong tenant boundaries, virtual machines, or dedicated hosts when justified by the risk assessment.
- Avoid sharing libraries, executable pages, and caches across trust domains when the platform provides appropriate separation mechanisms.
- Keep the operating system, runtime, hypervisor, and tools updated; exact mitigation is platform-dependent and should follow vendor guidance.
- Include side channels and co-location in the threat model, without treating an unprivileged process as an unobservable process.
- Monitor anomalous process behavior and workload density without collecting more sensitive content than necessary.
- Validate the boundary in a representative environment before claiming that the agent is isolated.
Limitations, counterpoints, and responses from the parties
The study is a v1 preprint, with no peer review or independent reproduction identified by the cutoff. The authors themselves built and evaluated the method. The set covers 11 models and three platforms, but it does not represent every provider, kernel, hypervisor, runtime, tool, or cloud topology.
The authors report responsible disclosure to the OpenAI and Anthropic security teams on September 21, 2026. The paper records no public response from the companies in the version consulted. The absence of a public response neither confirms nor refutes the finding. We also identified no vendor advisory, CVE, or independent corroboration that would justify turning the experiment into a claim of a widespread incident.
Practical decision
Do not use ‘the agent runs in a sandbox’ as a complete conclusion. Ask which resources remain shared, which adversary may be co-located, and what evidence proves the separation required by the risk. If the workflow handles secrets, proprietary code, financial operations, or regulated data, the physical and virtual boundary must be part of the same security contract as the tools.
Direct source
- ToolPulse: Inferring AI Agent Workflows via Microarchitectural Side Channels, arXiv:2611.25470v1, submitted on November 19, 2026.
Editorial and responsibility note
This text summarizes results reported by the authors and omits instructions that could facilitate abuse. Personal and professional views are not proven facts or guarantees of results. Trustyu did not reproduce the experiments and does not present the metrics as a general rate. The content is informational and does not replace technical, statistical, legal, regulatory, or security assessment. ToolPulse, Codex, Claude Code, and other marks belong to their respective owners; there is no sponsorship or endorsement. Tech Human and Trustyu operate commercially in digital-product development and governance. Research, structure, and writing received AI assistance; Fernando Parreiras authorized the topic, factual self-review, and publication on November 20, 2026, without additional independent human review. The paper remains under monitoring, and this brief may be corrected after peer review, new data, reproduction, an advisory, or responses from the parties.
Research cutoff: November 20, 2026. Editorial status: special Evidence Brief authorized for November 22, 2026 at 09:00 BRT.
Editorial and responsibility note
- Research cutoff
- Last review
- Recorded corrections
- No corrections recorded.
The cutoff above applies to the canonical claims. Additional sources and their access dates are identified in the article body.
This article combines cited sources, analysis, and the author's professional experience. Verifiable data and factual statements are linked to their respective sources. Interpretations, hypotheses, projections, recommendations, and opinions represent the author's professional point of view at the time of publication; they do not constitute proven facts, a promise of results, or legal, financial, or technical advice applicable to a specific case. Consult the original sources and qualified professionals before making decisions.
Claims and sources
R2-C3
Least privilege, isolated execution and supply-chain provenance are complementary controls; no substitutes for the others in an agentic system.
Limit: The exact control profile remains environment-specific and must be threat-modeled per product. This is a research-synthesis design input; it does not prove product adoption, operational maturity, independent attestation or outcome.
- OpenAI — OpenAI, Apache-2.0 repository; analysis-only excerpts
- OWASP Foundation — OWASP Foundation, CC-BY-SA-4.0 standard; analysis-only excerpts
- OpenSSF SLSA — OpenSSF SLSA, CC-BY-4.0 specification; analysis-only excerpts