Auditing the State Transfers That Execution Traces Miss in Long-Horizon LLM Agents
Key Facts
- While conventional execution traces log model prompts and tool calls, critical security and reliability failures occur during state transfers: a transfer can admit unchecked content (indirect prompt injections, unverified memory writes), rewrite state lossily (context compaction dropping user constraints, memory consolidation stripping source authority), or restore state incompletely (workspace rollbacks leaving aborted KV state active on the inference server).
- Computer Science > Cryptography and Security [Submitted on 17 Nov 2026] Title:Securing and Auditing Long-Horizon LLM Agents: State-Transfer Auditing and Controls in the Agent Harness View PDF HTML (experimental)Abstract:Long-horizon LLM agents across codebases, web browsers, and enterprise infrastructure rely on an execution harness to move state between the working context, session history, persistent memory, and external environment.
A paper submitted on 17 Nov 2026 to Computer Science > Cryptography and Security, “Securing and Auditing Long-Horizon LLM Agents: State-Transfer Auditing and Controls in the Agent Use,” examines long-horizon LLM agents operating across codebases, web browsers, and enterprise infrastructure. The paper describes an execution use that moves state between the working context, session history, persistent memory, and external environment. Conventional execution traces log model prompts and tool calls. The paper locates critical security and reliability failures in the state transfers between those locations.
Three Transfer Failures
The paper identifies three classes of state-transfer failure. A transfer can admit unchecked content, including indirect prompt injections and unverified memory writes. A transfer can rewrite state lossily: context compaction can drop user constraints, and memory consolidation can strip source authority. A transfer can restore state incompletely, with workspace rollbacks leaving aborted KV state active on the inference server.
Auditing and Controls
The title names state-transfer auditing and controls in the agent use, but the abstract provides no specific remediation or hunting guidance.
Sources & References
- [2611.23346] Securing and Auditing Long-Horizon LLM Agents: State-Transfer Auditing and Controls in the Agent Harness. Primary academic paper. https://arxiv.org/abs/2611.23346
- [2611.23346v1] Securing and Auditing Long-Horizon LLM Agents: State-Transfer Auditing and Controls in the Agent Harness. Primary academic paper. https://arxiv.org/abs/2611.23346v1
- Search | arXiv e-print repository. Primary academic paper. https://arxiv.org/search/cs?searchtype=author&query=Shah,+D
- Securing and Auditing Long-Horizon LLM Agents:State-Transfer Auditing and Controls in the Agent Harness. Primary academic paper. https://arxiv.org/html/2611.23346v1
- Primary academic paper referenced by arXiv Security (cs.CR). https://arxiv.org/pdf/2611.23346
SecurityXP delivers cybersecurity news, vulnerability analysis, data breach reports, and threat intelligence.
Security Digest
The most important cybersecurity stories, CVE alerts, and threat intelligence — one concise email. Unsubscribe anytime.
Related Articles
AI Approval Is Not Access Control: What the 97% Breach Statistic Actually Says
IBM's 2025 Cost of a Data Breach Report found that 97% of organizations reporting an AI-related breach lacked proper AI access controls. Approval is not the same as an access review.
Microsoft restricts employee Claude Fable 5 access over Anthropic data retention
Microsoft restricts employee access to Claude Fable 5 while legal reviews Anthropic's 30-day retention policy, which can retain flagged content for two years.
New EU AI Security Regulations for Organizations
The European Union has introduced comprehensive AI security regulations requiring organizations to implement security measures for AI systems. We break down the requirements, timelines, and compliance steps.
Essential Security Moves for Leaders in the Age of AI
As AI adoption accelerates in enterprises, security leaders must implement layered security strategies to mitigate evolving threats and protect complex digital architectures