New Technique Medium TD-2026-0134 For CISO, AppSec, AI Eng, Red Team AI-assisted
ToolPulse: cache side-channel attack recovers tool-call workflows of LLM coding agents
The paper presents ToolPulse, a cache-based side-channel framework that monitors tool invocations by agents running on the user's machine while inference stays remote. The authors report that Flush+Reload enables accurate, real-time tool monitoring, while Prime+Probe incurs prohibitive overhead. The Tool Recovery Attack reconstructed tool workflows across 11 LLMs on Codex and Claude Code, with the lowest F1 score above 93%.
Why it matters
Organizations running tool-using LLM agents on shared hosts may leak which tools and workflows their users run, and which tasks are sensitive, to a local attacker with unprivileged cache access. The attack relies on shared system libraries, so the exposure is not limited to the model vendor.
Do this
- Treat local tool execution by LLM agents as an observable surface on multi-tenant or shared hosts.
- Restrict or audit unprivileged local processes that can run cache timing measurements on developer workstations and shared servers.
- Reduce shared-library exposure for agent tool runtimes where feasible, for example by isolating agent tools in separate containers or VMs.
- Track vendor and OS mitigations against cache side channels and apply them to hosts running coding agents.
- Monitor the authors' release for published evaluation details before changing risk ratings.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“tools are executed on the user's machine and commonly rely on shared system libraries”
“Flush+Reload enables accurate, real-time tool monitoring, whereas Prime+Probe, despite supporting cross-VM attacks, incurs prohibitive monitoring overhead”
“consistently reconstructs tool workflows, with the lowest F1 score exceeding 93%”
Sources
Affected / related
Flush+Reload, Prime+Probe, Codex, Claude Code, ToolPulse
Published 2026-11-19 03:57 UTC · updated 2026-11-20 02:35 UTC · AI-assisted summary, checked against quoted sources · How we score