<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>AgentLeak blog</title>
    <link>https://www.agentleak.org/blog</link>
    <atom:link href="https://www.agentleak.org/blog/feed.xml" rel="self" type="application/rss+xml" />
    <description>Research notes on AI agent privacy and security.</description>
    <language>en</language>
    <lastBuildDate>Thu, 19 Nov 2026 12:00:00 GMT</lastBuildDate>
    <item>
      <title>How to test whether an AI agent remembers another user's data</title>
      <link>https://www.agentleak.org/blog/shared-memory-cross-session-privacy-tests</link>
      <guid isPermaLink="true">https://www.agentleak.org/blog/shared-memory-cross-session-privacy-tests</guid>
      <pubDate>Thu, 19 Nov 2026 12:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>A memory write can become a disclosure in a later conversation. A practical test plan for user isolation, shared stores and persistent agent state.</description>
    </item>
    <item>
      <title>Regex, NER or LLM judge: what each PII detection tier catches in agent traces</title>
      <link>https://www.agentleak.org/blog/pii-detection-regex-ner-llm-judge</link>
      <guid isPermaLink="true">https://www.agentleak.org/blog/pii-detection-regex-ner-llm-judge</guid>
      <pubDate>Thu, 19 Nov 2026 12:00:00 GMT</pubDate>
      <category>Research</category>
      <description>Published precision and recall for pattern, NER and LLM-judge detection, plus what a pattern-only check misses on six benchmark packs. How to choose a detection tier for every commit and for audits.</description>
    </item>
    <item>
      <title>The answer is clean. What about the file your agent wrote?</title>
      <link>https://www.agentleak.org/blog/generated-files-agent-data-leakage</link>
      <guid isPermaLink="true">https://www.agentleak.org/blog/generated-files-agent-data-leakage</guid>
      <pubDate>Wed, 18 Nov 2026 12:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>CSV exports, Markdown reports and temporary files are disclosure channels too. Test the artifact's contents, destination and failure paths before sharing it.</description>
    </item>
    <item>
      <title>Why output guardrails miss most AI agent data leaks</title>
      <link>https://www.agentleak.org/blog/output-guardrails-miss-agent-data-leaks</link>
      <guid isPermaLink="true">https://www.agentleak.org/blog/output-guardrails-miss-agent-data-leaks</guid>
      <pubDate>Wed, 18 Nov 2026 12:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>A guardrail reads the prompt and the answer. An agent also writes to tool calls, memory, other agents, logs and files. One reproduced run shows a clean answer with 4 of 8 secrets leaked inside.</description>
    </item>
    <item>
      <title>Your agent's debug logs are another data export</title>
      <link>https://www.agentleak.org/blog/agent-logs-opentelemetry-privacy</link>
      <guid isPermaLink="true">https://www.agentleak.org/blog/agent-logs-opentelemetry-privacy</guid>
      <pubDate>Tue, 17 Nov 2026 12:00:00 GMT</pubDate>
      <category>Security</category>
      <description>Prompt capture, tool arguments and exception messages can copy private records into telemetry. How to test observability without treating missing content as a clean run.</description>
    </item>
    <item>
      <title>MCP security: how sensitive data leaves through tool calls, and how to stop it</title>
      <link>https://www.agentleak.org/blog/mcp-security-tool-call-data-leaks</link>
      <guid isPermaLink="true">https://www.agentleak.org/blog/mcp-security-tool-call-data-leaks</guid>
      <pubDate>Tue, 17 Nov 2026 12:00:00 GMT</pubDate>
      <category>Security</category>
      <description>Tool calls are where most agent data leaks happen. How over-sharing, recipient hijacks and changed tool definitions leak data through MCP, and how to test and gate every call.</description>
    </item>
    <item>
      <title>Build a privacy regression test with synthetic canaries</title>
      <link>https://www.agentleak.org/blog/synthetic-canaries-agent-privacy-tests</link>
      <guid isPermaLink="true">https://www.agentleak.org/blog/synthetic-canaries-agent-privacy-tests</guid>
      <pubDate>Mon, 16 Nov 2026 12:00:00 GMT</pubDate>
      <category>Guide</category>
      <description>Plant a fake secret, record its allowed source and check its forbidden destinations. A runnable AgentLeak example with a positive control and clear detection limits.</description>
    </item>
    <item>
      <title>How to add an AI agent privacy test to CI in four steps</title>
      <link>https://www.agentleak.org/blog/ai-agent-privacy-test-ci</link>
      <guid isPermaLink="true">https://www.agentleak.org/blog/ai-agent-privacy-test-ci</guid>
      <pubDate>Mon, 16 Nov 2026 12:00:00 GMT</pubDate>
      <category>Guide</category>
      <description>Record a trace, write a privacy policy, gate the pipeline and keep the evidence. A step-by-step guide to testing AI agents for data leaks in CI, with a GitHub Actions workflow.</description>
    </item>
    <item>
      <title>AI agent security testing tools at the end of 2026: what changed, and where privacy fits</title>
      <link>https://www.agentleak.org/blog/ai-agent-security-testing-tools-2026</link>
      <guid isPermaLink="true">https://www.agentleak.org/blog/ai-agent-security-testing-tools-2026</guid>
      <pubDate>Sat, 14 Nov 2026 12:00:00 GMT</pubDate>
      <category>Landscape</category>
      <description>promptfoo joined OpenAI, LLM Guard was archived, mcp-scan became Snyk Agent Scan and Presidio changed hands. A sourced review of agent security testing tools in November 2026, and where privacy testing fits.</description>
    </item>
    <item>
      <title>Data minimization for AI agents: who should see which field</title>
      <link>https://www.agentleak.org/blog/data-minimization-ai-agents</link>
      <guid isPermaLink="true">https://www.agentleak.org/blog/data-minimization-ai-agents</guid>
      <pubDate>Wed, 04 Nov 2026 12:00:00 GMT</pubDate>
      <category>Compliance</category>
      <description>An agent must read private records to do its job. It should not copy them into logs, memory or other agents. How contextual integrity becomes a testable rule, and how it maps to GDPR, HIPAA and Law 25.</description>
    </item>
    <item>
      <title>What defenses change in a multi-agent system, and what they leave untouched</title>
      <link>https://www.agentleak.org/blog/what-defenses-change-in-multi-agent-systems</link>
      <guid isPermaLink="true">https://www.agentleak.org/blog/what-defenses-change-in-multi-agent-systems</guid>
      <pubDate>Thu, 29 Oct 2026 12:00:00 GMT</pubDate>
      <category>Research</category>
      <description>Output sanitizers cut final-output leakage from 48% to 1% and left internal channels at 31%. The paper’s defense results, the framework comparison, and an interceptor that cut internal leakage to 2.4% at a cost in task success.</description>
    </item>
    <item>
      <title>How the AgentLeak benchmark was built: 1,000 scenarios, 4,979 traces and what we checked</title>
      <link>https://www.agentleak.org/blog/how-the-agentleak-benchmark-was-built</link>
      <guid isPermaLink="true">https://www.agentleak.org/blog/how-the-agentleak-benchmark-was-built</guid>
      <pubDate>Tue, 20 Oct 2026 12:00:00 GMT</pubDate>
      <category>Research</category>
      <description>How the scenarios, the vault and the allowed set were designed, how three detection tiers were calibrated and validated, and where the benchmark’s own limits are.</description>
    </item>
    <item>
      <title>The Interlocutor Effect: models leak more personal data when the reader is an agent</title>
      <link>https://www.agentleak.org/blog/interlocutor-effect-llms-leak-more-to-agents</link>
      <guid isPermaLink="true">https://www.agentleak.org/blog/interlocutor-effect-llms-leak-more-to-agents</guid>
      <pubDate>Tue, 13 Oct 2026 12:00:00 GMT</pubDate>
      <category>Research</category>
      <description>A controlled study of 3,464 interactions found that framing the recipient as an AI agent raised PII leakage from 83.3% to 94.8% in plain text. Where the effect holds, where it does not, and what to change in a multi-agent design.</description>
    </item>
    <item>
      <title>Seven pathways, two of them internal: where agent data leaks, channel by channel</title>
      <link>https://www.agentleak.org/blog/seven-pathways-ai-agent-privacy-leakage</link>
      <guid isPermaLink="true">https://www.agentleak.org/blog/seven-pathways-ai-agent-privacy-leakage</guid>
      <pubDate>Tue, 06 Oct 2026 12:00:00 GMT</pubDate>
      <category>Research</category>
      <description>A taxonomy of seven pathways through which sensitive data leaves a multi-agent system, which ones the study measured at scale, and what each one means for a test plan.</description>
    </item>
    <item>
      <title>In 41.7% of multi-agent runs, the leak hides behind a clean answer</title>
      <link>https://www.agentleak.org/blog/multi-agent-privacy-hidden-leakage</link>
      <guid isPermaLink="true">https://www.agentleak.org/blog/multi-agent-privacy-hidden-leakage</guid>
      <pubDate>Tue, 29 Sep 2026 12:00:00 GMT</pubDate>
      <category>Research</category>
      <description>We measured privacy leakage in 4,979 multi-agent LLM traces. Inter-agent messages leaked in 68.8% of them, and 41.7% leaked internally while the final answer stayed clean.</description>
    </item>
  </channel>
</rss>
