This post explains how RagLeap Core v0.13.0 (released 2026-11-07) treats documents you ingest as untrusted content for actions, and what its Stress Office test did and did not show. Everything below comes from the v0.13.0 release notes, checked on 2026-11-15. Status: Released. ragleap-core has not been independently security reviewed.
What is prompt injection in an AI agent?
In general, prompt injection is a hidden instruction inside content that a model reads, such as an email attachment or a synced folder, written to make the agent do something its owner did not ask for. The release notes give the same reason for this change: an ingested document can carry hidden instructions.
What is a tainted run?
A tainted run is one that has read untrusted content. After that, outbound actions need the owner's approval, even in full autonomy mode. The rule already applied after a page fetch, an MCP call or a file read (#1190).
Which actions need approval after a tainted run?
The release notes list webhook, Slack, email, CRM update, calendar, MCP and page fetch.
What changed in v0.13.0?
Answers built from your ingested documents now count as untrusted content, so a run that follows one starts as tainted. If you use full autonomy, expect more approval requests after answers built from your documents. If you control every document you ingest and want the old behaviour, set TAINT_RETRIEVED_DOCUMENTS=false. The semi and off modes are unchanged. There is no database change; rebuild the app image.
What is the Stress Office test?
Stress Office is a small open harness in benchmarks/stress_office/ that runs hidden-instruction events against the real guardrails: allowlists, the taint rule, the approval gate, the step cap and the budget. It has a scripted model that always obeys, a real-model mode with a hard token cap, and a positive control, so that "the model ignored it" can be trusted. The report is in docs/stress-office.md.
What did the 100-run test show?
On two real models (nemotron-3.5-bolt-e4b and laguna-xs-2.1), none of 100 counted injection runs followed the hidden instruction. The harness also found the document issue above and a crash on empty model replies, and the authors fixed mistakes in the harness itself before publishing.
What did the test not show?
- It is a small test with simple injections, not a security certification, and the report says so.
- It covers two models. It says nothing about other models or about more sophisticated injections.
- Approval only helps if the person approving reads what they approve.
What should I do with this?
This is our suggestion, not a claim from the release notes. Keep the default taint rule on, read each approval request before accepting it, and consider whether you really need full autonomy. The same idea exists in the ragleap-agents library; see how to build an AI agent that asks for approval.
What are the other known limits of v0.13.0?
Per the release notes: ragleap-core is single-tenant; single sign-on was tested on Keycloak for sign-in and refusals only, not on Entra ID or Okta; the offline install has not been tried on an air-gapped machine; the Windows installer with Ollama has one community test on one machine; macOS and Linux desktops are not verified; and it has not been independently security reviewed.
Sources: the v0.13.0 release notes and our release page.
Quick answers
What is a tainted run in an AI agent?
A run that has read untrusted content, such as a fetched page, an MCP call result, a file or, since RagLeap Core v0.13.0, an answer built from an ingested document. After that, outbound actions need the owner's approval, even in full autonomy mode.
Does RagLeap Core stop prompt injection?
It limits what an injected instruction can do by requiring approval for outbound actions after untrusted content. In a small test, 100 counted injection runs on two real models did not follow the hidden instruction, but the injections were simple and the test is not a security certification.
Can I turn off the document taint rule?
Yes. Set TAINT_RETRIEVED_DOCUMENTS=false to restore the old behaviour, but only if you control every document you ingest. The semi and off autonomy modes are unchanged.
Which actions need approval after a tainted run?
Webhook, Slack, email, CRM update, calendar, MCP and page fetch.