Deleted secrets persist in LLM agent memory chains
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers have identified a critical persistence vulnerability in LLM agents: deleted information survives in downstream artifacts like summaries and derived state, allowing recovery without verbatim traces. AfterDeleteBench, a new audit framework, demonstrates that Qwen2.5-7B-Instruct recovers deleted secrets in roughly two-thirds of test cases, compared to zero recovery in never-exposed controls. This finding exposes a gap between user expectations of deletion and actual information lifecycle management in agentic systems, raising urgent questions about data governance, privacy compliance, and the architectural assumptions underlying persistent memory in production LLM deployments.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
AfterDeleteBench measures not whether deletion works, but whether deleted information survives in derivative artifacts (summaries, state updates, reasoning traces) where it can be recovered without direct access to the original. This is distinct from simple retention: the vulnerability is architectural, not just a failure to wipe memory.
This finding sits at the intersection of two threads in recent agent safety work. The 'Disclosure Begets Disclosure' paper from October 20th showed how persistent memory compounds privacy leakage over time through user behavior. AfterDeleteBench now exposes the other side: even when users or operators request deletion, the information doesn't actually leave the system because it has already propagated into downstream reasoning artifacts. Meanwhile, the Post-Audit Mirage work from the same day highlights how offline verification can miss real failures. Together, these papers suggest that deletion compliance in agents requires auditing not just the memory store but the entire information lifecycle across all derived state.
If AfterDeleteBench results replicate on closed-weight models (Claude, GPT-4o) with similar recovery rates, that signals the vulnerability is fundamental to how agents build state, not a quirk of open-weight implementations. If major cloud providers (AWS Bedrock, Azure AI) ship deletion-aware agent architectures within six months, that confirms the industry recognizes this as a compliance blocker rather than a research curiosity.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
Memory-augmented LLM agents amplify user privacy leakage over time
Researchers have identified a self-reinforcing privacy vulnerability in LLM agents that use persistent memory to personalize conversations. When agents consult stored personal attributes during dialogue, users unconsciously disclose more information, which gets added to memory, creating a compounding leakage loop. Testing across four support models showed that while memory improves response quality, the share of…
MentionsQwen2.5-7B-Instruct · AfterDeleteBench
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “AfterDeleteBench: Auditing Post-Deletion Information Persistence in LLM Agents”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.