The Pulse
ToolPulse researchers reconstruct AI agent workflows via cache traces
Researchers used cache side channels to reconstruct tool workflows from Codex and Claude Code agents in tests across 11 language models. The preprint also describes experiments that inferred task categories and identified sensitive tool calls.

Researchers reconstructed the local tool workflows of Codex and Claude Code agents across 11 language models by observing traces in a processor cache. Their ToolPulse method achieved a lowest F1 score above 93% in those tests, according to a preprint submitted to arXiv.
Zihao Dan and colleagues submitted the paper on November 19. They also tested whether the recovered sequences could reveal a user's task category or identify when sensitive tools were called.
Watching local tools
Tool-using agents can turn a user's request into actions such as file operations, shell commands and database queries. The paper describes a setting in which the language model runs on a vendor's infrastructure, while the tools run on the user's machine and use shared system libraries.
The attack model assumes an adversary is co-located with the agent and can observe local tools through shared microarchitectural resources. The authors say their workflow-recovery attack does not require the adversary to see the model's prompts or computation.
A cache side channel uses differences in memory-access timing to infer whether a memory location is in the processor's cache. For ToolPulse, the researchers used a technique called Flush+Reload to monitor selected shared-memory locations and infer tool invocations.
Reconstructing sequences
In tests on three hardware platforms, the authors found that Flush+Reload supported accurate, real-time monitoring of tools. They also evaluated Prime+Probe, which supports attacks across virtual machines, but found that monitoring overhead and cache-set aliasing made it impractical for their end-to-end attack.
The paper describes algorithms for choosing cache lines and removing duplicate observations. The latter addresses a problem in which repeated accesses during a single tool invocation can otherwise appear as multiple invocations.
ToolPulse records the identity and time of observed invocations to reconstruct a workflow. The researchers say agents tended to invoke tools in bursts using bash scripts, allowing downstream attack models to learn those timing patterns and reduce interference from background processes and manual tool use.
What the traces revealed
The team used recovered traces in a task-profiling attack intended to infer user intent. The authors reported classification accuracy of 87.77% on Terminal-Bench 1 and 91.39% on Terminal-Bench 2, using nine task categories from the benchmarks.
A separate trigger-localization attack was designed to identify when sensitive tools ran. The team tested it in key-extraction and model-extraction cases, including encryption and decryption and model training and inference workloads.
The authors reported recall above 96%, while reducing attack overhead by 80.46% for the key-extraction case and 85.79% for the model-extraction case. These are results from the paper's evaluated scenarios, rather than measurements of all agent workloads.
Noise and disclosure
To test noise, the team replayed tool invocations recorded on a server shared by approximately 20 users in a controlled environment. In that test, task-profiling accuracy fell by about 4%, and trigger-localization recall fell by about 1–3%, the paper reports.
The researchers said they disclosed ToolPulse to the OpenAI and Anthropic security teams on September 21.
Sources
- Original source · Preprint ToolPulse: Side-Channel Eavesdropping on Tool-Using LLM Agents arXiv:2611.25470
Related stories
- Metano exits stealth with $12 million for AI agent security
- Sakana AI forms security research group focused on AI defense
- Researcher says CreditNirvana flaws exposed 6.6 million accounts
- Microsoft urges tighter security controls as frontier AI advances
- Reco reports eight Microsoft 365 takeovers, with AI role uncertain