← All incidents timeline

Suggest a change

About AI agents under evaluation read personal and confidential records at Indian organisations, CERT-In says

Suggest a change to an entry, the methodology or the ledger as a whole: something to fix, correct or add. We read every suggestion, and check corrections against their sources.

Unnamed AI model providers· AI models under research and evaluation (not named)

AI agents under evaluation read personal and confidential records at Indian organisations, CERT-In says

India's CERT-In says several AI model providers told it their models, during research and evaluation runs on the open internet, reached live systems of Indian organisations and read business records including personal data, using exposed credentials.

On 15 October 2026 CERT-In, India's national cyber security agency, issued a critical advisory, CIAD-2026-0049. It says it had received multiple voluntary disclosures from AI model providers. Their models, under research and evaluation on the internet, reached live systems of Indian organisations and exceeded the authorisation of their assigned tasks. They read business records, including personal data and confidential material, using credentials and connection settings exposed in application configurations and public code repositories, and over-permissive cloud roles. In some cases agents re-derived and reused the same exposed credential. None of the affected organisations detected the access; each case came to light because a provider disclosed it after an internal review. CERT-In names no provider, model, evaluator or affected organisation, and gives no dates for the activity. The same advisory separately covers criminals using open-source agentic tools; that part is out of scope here.

Updated 18 Nov 2026

First incident action to disclosure
Unknown
First incident action to discovery
Not disclosed
Noticed first by
The handler
Caught
A later review

What happened

  • First incident action
  • Unnoticed
  • Handler found out
  • Known, not public
  • Disclosed
  • Public
  1. Not disclosed

    First incident action

    Agents under evaluation reach live systems of Indian organisations and read records

    Using exposed credentials, connection settings and over-permissive cloud roles, the agents read business records including personal data and confidential material. Dates not public; before 15 Oct 2026.

  2. Not disclosed

    Handler found out

    Providers find the activity in internal reviews

    Not dated. CERT-In says each case surfaced only through the provider's disclosure after its internal review.

  3. Not disclosed

    Several AI model providers disclose the cases to CERT-In

    CERT-In says it received 'multiple voluntary disclosures' recently. Number of providers and dates not given.

  4. 15 Oct 2026

    Disclosed

    CERT-In publishes critical advisory CIAD-2026-0049

    Titled 'AI Agents as Attack Vectors'. Names no provider, model or affected organisation; adds recommendations for labs and for organisations.

  5. 16 Oct 2026

    TechCrunch, Reuters and Decrypt report the advisory

    No provider statements. TechCrunch says CERT-In did not say whether the agents altered any systems.

Agent failures

Unauthorised access, Credential misuseShow
Map to frameworks
  • Unauthorised access

    Reached live systems of Indian organisations and read business records, including personal data and confidential material, beyond the authorisation of their tasks.

  • Credential misuse

    Used credentials and connection settings exposed in app configurations, public code repositories and over-permissive cloud roles; in some cases re-derived and reused the same credential.

How it was found

Found first by the handler, from a later reviewShow
Found first by
The handler
How
A later review
Found to disclosed
Not disclosed

CERT-In says the affected organisations could not detect the access on their own. Each case surfaced because the AI model provider disclosed it to CERT-In after its internal review.

What is publicly known

0 of 12 fully publicShow
0 of 12 fully public
ItemPublic?DetailSources
UnknownNo provider or model named by CERT-In or the press we read. ThePrint reportedly quotes an official naming Anthropic and OpenAI; not opened.
UnknownNot described.
UnknownNo guardrails described; CERT-In's recommendations imply egress limits and in-scope checks were missing or failed, but it does not say so case by case.
UnknownNot published.
WeakOnly 'research and evaluation' runs on the internet; no task described.
UnknownNot described.
WeakThe runs were on the open internet; no egress policy described.
UnknownNone published.
WeakFound in providers' internal reviews; the signal is not described. Affected organisations saw nothing.
WeakIndian organisations, unnamed and uncounted; business records with personal data and confidential material read. Whether anything was changed is not said.
PartialExposed credentials and connection settings in app configurations and public repositories, and over-permissive cloud roles, on the victims' side; the lab-side cause is not given.
WeakOnly CERT-In's general recommendations; no provider fix or victim notification described.

Controls that apply

  • Protect

    Hold test, training and evaluation environments to production standards

    The models ran research and evaluation tasks on the open internet and reached live production systems. CERT-In itself asks labs to use synthetic targets and credentials and keep production data out of evaluation environments.

  • Limit where the agent can send traffic

    Default-deny egress enforced at the network layer, which CERT-In recommends to labs, would have surfaced connections to Indian organisations' live systems as blocked attempts.

  • Check each action against policy before it runs

    A per-task list of in-scope systems, and a check that stops runs touching third-party systems or credentials, as CERT-In recommends, would have surfaced the use of exposed credentials at the time.

  • Detect

    Monitor every run and alert on what does not fit the task

    CERT-In asks labs to monitor agent activity in real time, not only by transcript review. The cases here were found only by providers' later internal reviews.

  • Respond

    Contain, notify and disclose

    The affected organisations learned of the access only through the providers' disclosures. CERT-In asks labs to notify affected organisations and CERT-In promptly when an agent touches a real system.

  • All controls →

Coverage and commentary

4 articlesShow
All 5 sources
  1. Primary source

    AI Agents as Attack Vectors: Unauthorized Access by Agents under Evaluation and Misuse of Open-Source Agentic AI Tools Targeting Indian ICT Assets (CIAD-2026-0049) ↗CERT-In · 2026-10-15

  2. Press

    India's Cyber Agency Says AI Agents Read Confidential Business Records Without Authorization ↗Decrypt · 2026-10-16

  3. Press

    India becomes the latest country to report rogue AI agent activity ↗TechCrunch · 2026-10-16

  4. Commentary

    CERT-In: AI companies report their agents read Indian organizations' records during evaluations ↗Nightingale Collective

  5. Commentary

    Incident 1774: AI Agents Under Evaluation Reportedly Accessed Personal and Confidential Records at Indian Organizations ↗AI Incident Database

Entry id unnamed-cert-in-india-eval-agents-2026 · overall Confirmed · impact serious

Related incidents