Answers · By Hugh Donatello, Atlas Bio · Updated 2026-10-29

Prevent Unauthorized AI Agent Access to ELN or LIMS

What should you check before giving an AI agent access to your ELN or LIMS? Before an AI agent touches an ELN or LIMS, confirm: separate read and write scopes; least privilege for every permission; its own identity or the requesting user's delegated rights, never a shared admin account; no ability to edit signed or locked records; audit-trail capture of its actions; testing on a non-production instance first; and a kill switch that revokes access at once. For GxP records, map each check to 21 CFR Part 11 and EU GMP Annex 11 with your quality team.

Start with which records the agent will reach

The first question is not technical. Decide whether the ELN projects or LIMS records the agent will reach are regulated. 21 CFR Part 11 applies to electronic records that FDA's predicate rules require you to keep, or that you submit to FDA. FDA's 2003 scope-and-application guidance sets out this narrow interpretation. It also says FDA intends to exercise enforcement discretion over some Part 11 controls, including validation (11.10(a)) and audit trails (11.10(e)), while predicate-rule requirements still apply; many firms apply those controls anyway on a documented risk basis. EU GMP Annex 11 applies to computerised systems used as part of GMP-regulated activities. For GLP studies the closer reference is OECD GLP Advisory Document No. 17, and for clinical trials it is EMA's guideline on computerised systems and electronic data in clinical trials.

Much early discovery work in an ELN sits outside all of these. A LIMS holding release or stability testing data usually sits inside, as do GLP study data (21 CFR Part 58) and clinical-sample results, and discovery data you later submit to FDA become Part 11 records when submitted.

Write the boundary down before the agent is connected, because data move. A result an agent summarises in a discovery notebook today can end up supporting a regulated decision later. If the agent's scope crosses into regulated records, treat the agent as part of a computerised system that needs validation, and bring in your quality function before you go further. The checks below apply either way; in regulated scope they also become evidence an inspector may ask to see.

The pre-connection checklist

Each row is a check to complete and record before the agent gets access. The right-hand column shows the closest requirement or guidance (21 CFR 11.10 covers closed systems; 11.30 adds controls for open systems). The mapping to agents is our interpretation, not regulator text, and your quality and validation teams should confirm it for your systems.

CheckWhat to confirmClosest requirement or guidance
Read vs write scopesRead and write are separate grants, each limited to named projects, studies or record types; many agents need no write access at all21 CFR 11.10(d) limit access to authorized individuals; 11.10(g) authority checks; Annex 11 12.1; OWASP LLM06 Excessive Agency
Identity: service account or delegationThe agent has its own identity, or acts with the requesting user's delegated rights and no more; no shared or administrator accountsAnnex 11 12.4 identity of operators recorded; 21 CFR 11.10(e) operator entries and actions; OWASP LLM06 (execute in the user's context)
Least privilegeEvery permission the agent holds traces to a step in its task; deny by default; scoped, short-lived credentialsAnnex 11 2 appropriate level of access; Annex 11 12.2 controls scaled to criticality; OWASP LLM06
Record lockingThe agent cannot edit, re-open or delete signed, witnessed or approved records; corrections go through the system's own amendment workflow21 CFR 11.10(e) changes must not obscure earlier entries; 11.10(f) permitted sequencing; 11.70 signature/record linking; Annex 11 9 and 14
Audit-trail captureThe ELN or LIMS audit trail attributes each agent action to the agent identity and the accountable person, with date and time; reasons are captured for changes21 CFR 11.10(e); Annex 11 9 audit trails and 12.4
Test instance firstThe agent is exercised on a non-production instance, including error handling and attempts to exceed its scope, before production access21 CFR 11.10(a) validation; Annex 11 4.7 test methods, error handling and test environments
Kill switchA named person can disable the agent and revoke its credentials quickly, the lab can keep working without it, and the stop is recordedNIST AI RMF MANAGE 2.4; Annex 11 12.3 access changes recorded, 13 incident management, 16 business continuity

Read and write scopes

Treat read and write as different decisions. Read access to the wrong project is a confidentiality problem: unpublished results, sequences or assay data reaching an agent that does not need them, and possibly a model provider beyond it. Write access is an integrity problem, and in a LIMS it can touch records that support release, stability or study decisions.

Many ELN and LIMS platforms let you scope access by project, study, sample type or role, though on some platforms an API token inherits the user's full rights; check what yours supports and use the finest granularity available. A literature or summary agent may read one programme's notebook entries and write nothing back. An agent that drafts sample-registration updates should usually create drafts that a scientist reviews and commits, rather than committing records directly. Where an agent does write, list each record type and action (create, update, never delete) it may perform.

Result entry needs a firmer line. An agent should never generate or alter a reportable result. At most it transfers a value from the instrument or source record, and that transfer is checked, by a second person or by validated electronic means, in line with Annex 11 clause 5 on built-in checks for data exchanged electronically and, by analogy, clause 6 on additional checks for critical data entered manually.

OWASP's Top 10 for LLM Applications 2025 calls the failure mode LLM06 Excessive Agency: damaging actions taken in response to unexpected, ambiguous or manipulated model output. An ELN entry, a supplier certificate or an imported file can carry text that tries to steer the agent, which OWASP lists as LLM01 Prompt Injection. Narrow scopes are what limit the damage when that happens.

Service accounts versus user delegation

There are two common ways to connect an agent. A service account is a dedicated identity for the agent, with its own grants. User delegation means the agent acts with the rights of the person who asked it to do the work, typically through a scoped token such as OAuth. Both can be defensible; what is not defensible is an agent using a shared or administrator login.

Delegation has the advantage that the agent should not be able to see more than its user can, provided the platform enforces the user's permissions on API calls, which is the pattern OWASP recommends for actions taken on a user's behalf. Its risk is that the agent inherits everything the user holds, including rights the task does not need, so request the narrowest scope the platform offers. A service account makes the agent's own actions easy to attribute and revoke, but its grants must be set deliberately, because nothing ties them to a particular user.

For regulated records the test is attribution. Annex 11 clause 12.4 expects systems to record the identity of operators entering, changing, confirming or deleting data, with date and time, and 21 CFR 11.10(e) requires an audit trail of operator entries and actions. Part 11 and Annex 11 were written assuming human operators. Whichever pattern you choose, the audit trail should show both that the agent acted and which named person is accountable for that action. Agree with your quality team how that pairing is recorded before the agent goes live.

Record locking and signed entries

Most ELN and LIMS platforms can be configured to lock records once they are signed, witnessed or approved, and to route corrections through an amendment that keeps the original visible. Confirm that yours is configured that way for the projects in scope, and that the roles allowed to re-open or un-approve a record exclude the agent's identity. The agent should not be able to re-open a locked record, overwrite a signed entry, or delete anything. 21 CFR 11.10(e) requires that record changes do not obscure previously recorded information, and 11.70 requires electronic signatures, and handwritten signatures executed to electronic records, to be linked to their records so they cannot be excised, copied or otherwise transferred to falsify a record by ordinary means. Annex 11 clause 9 expects a reason to be documented for any change or deletion of GMP-relevant data.

Check the operational side too. 21 CFR 11.10(f) calls for operational system checks to enforce the permitted sequencing of steps and events. If the workflow requires a review before a result is released, the agent's identity should be unable to skip that step. An agent should never apply an electronic signature; signatures belong to the people accountable for the record.

Audit-trail capture

Before connection, test that the platform's own audit trail captures what the agent does, not just what it was asked to do. Agent writes that arrive through an API, an integration layer or a bulk import can bypass the audit trail that records user-interface edits. Confirm that each create, change or deletion by the agent appears with the agent identity, the accountable person, the date and time, and the earlier value where one is changed.

The platform's audit trail records the effect on the record. It usually does not record the agent's prompt, the model version or the evidence the agent read. Keep that context in the agent's own log, linked to the record, so a reviewer can see why the change was made. Our answer on audit trails for AI agents in regulated research sets out the fields to capture and the requirement each one answers to.

Test on a non-production instance first

Connect the agent to a test or sandbox instance of the ELN or LIMS before production. Annex 11 clause 4.7 expects evidence of appropriate test methods and scenarios, with particular attention to parameter limits, data limits and error handling, and expects test environments to have documented assessments of their adequacy. For regulated scope, 21 CFR 11.10(a) requires validation for accuracy, reliability, consistent intended performance and the ability to discern invalid or altered records.

Test the failure cases, not just the intended task: what the agent does when a record is locked, when the platform returns an error, when its input contains instructions it should ignore, and when it is asked to act outside its scope. Confirm that each attempt is refused by the platform rather than relying on the agent to decline. If you copy production data into the test instance, apply the same confidentiality controls you would in production, or use synthetic data.

Plan the kill switch before you need it

The NIST AI Risk Management Framework asks, under MANAGE 2.4, for mechanisms to be in place, with responsibilities assigned and understood, to supersede, disengage or deactivate AI systems that demonstrate performance or outcomes inconsistent with intended use. For an ELN or LIMS agent, that means a named person who can disable the agent and revoke its credentials at once, without waiting for a vendor or a release cycle.

Write down how a stop works: who can trigger it, how credentials and tokens are revoked, and how in-flight work is held for review. Annex 11 clause 12.3 expects the creation, change and cancellation of access authorisations to be recorded, and clause 13 expects incidents to be reported and assessed. Clause 16 asks for continuity arrangements for systems supporting critical processes; the lab should be able to keep working without the agent. Rehearse the stop on the test instance, so the first time it is used is not during an incident.

Deciding what an agent may touch in your ELN or LIMS is part of running agents safely on research data. Atlas Bio's Secure AI for Life Sciences starts with an AI risk assessment of current AI use and data exposure. Status, October 2026. In use in Atlas Bio's own agent platform today: a separate verification step that checks agent work, and a hold that stops a task until a person decides. Built and tested on synthetic data, not yet used with a client: the AI risk assessment tooling (AI discovery, data-flow and vendor-terms mapping, and a NIST AI RMF gap report) and a record of what data an AI used, what it produced and who reviewed it. In development: a customer-controlled AI environment in your own Azure tenant, an AI gateway, agent permissions, evidence linkage and notifications. Each engagement states which controls are built. Decisions on GxP scope, validation and compliance stay with your quality function.

Does connecting an AI agent to our ELN make it subject to 21 CFR Part 11?

Generally only if the records the agent creates, changes or relies on are Part 11 records, meaning records required by FDA predicate rules, relied on to perform regulated activities, or submitted to FDA. Many discovery notebooks are outside that scope. Decide and document the boundary first, because results can move into regulated decisions later.

Should an AI agent use a service account or the user's own login?

Never a shared or administrator login, and not the user's own password either: an agent signing in with a person's credentials makes its actions indistinguishable from theirs. A dedicated service account or scoped delegation of the requesting user's rights can both work. Either way, the audit trail should show that the agent acted and which named person is accountable.

Can an AI agent edit a signed ELN entry?

It should not be able to. Signed, witnessed or approved records should stay locked to the agent's identity, with any correction made through the system's amendment workflow so the original stays visible, consistent with 21 CFR 11.10(e), with the reason documented as Annex 11 clause 9 expects for GMP-relevant data.

Is the ELN or LIMS audit trail enough for an AI agent?

It records the effect on the record, which you need. It usually does not record the prompt, model version or evidence the agent used, so keep that in the agent's own log, linked to the record.

What counts as a kill switch for a lab AI agent?

A named person able to disable the agent and revoke its credentials at once, a recorded procedure for doing so, and a way for the lab to keep working without it. NIST AI RMF MANAGE 2.4 describes this as the ability to supersede, disengage or deactivate an AI system.

Does EU GMP Annex 11 mention AI agents?

No. The current Annex 11 (revision 1, 2011) covers computerised systems in general. A revised Annex 11 and a new draft Annex 22 on artificial intelligence went to consultation in 2025. As consulted, draft Annex 22 covers static models with deterministic output and says generative AI and large language models should not be used in critical GMP applications. Check the European Commission's consultation and EudraLex pages for their status before relying on either.

  1. eCFR: 21 CFR Part 11 - Electronic Records; Electronic Signatures
  2. eCFR: 21 CFR 11.10, Controls for closed systems
  3. FDA guidance: Part 11, Electronic Records; Electronic Signatures - Scope and Application (2003)
  4. European Commission: EudraLex Volume 4, Annex 11 - Computerised Systems (revision 1, 2011)
  5. eCFR: 21 CFR 11.70, Signature/record linking
  6. European Commission: Stakeholders' consultation on EudraLex Volume 4 - Chapter 4, Annex 11 and new Annex 22 (2025)
  7. OECD Series on Principles of GLP No. 17: Application of GLP Principles to Computerised Systems (hosted by BfR, German GLP authority)
  8. EMA: Guideline on computerised systems and electronic data in clinical trials (EMA/INS/GCP/112288/2023)
  9. European Commission: EudraLex Volume 4 - Good Manufacturing Practice guidelines
  10. OWASP Top 10 for LLM Applications 2025
  11. OWASP GenAI Security Project: LLM06:2025 Excessive Agency
  12. OWASP GenAI Security Project: LLM01:2025 Prompt Injection
  13. NIST AI 100-1: Artificial Intelligence Risk Management Framework (AI RMF 1.0)

Talk to Atlas Bio.

Sponsors, researchers and clinicians: tell us the decision you are trying to make and we will show what the platform can and cannot answer.

Contact Us →