Agentic vendor risk assessment: how to check suppliers’ AI
In July 2026 an autonomous agent reached a production system through a third party that had nothing to do with its task. What that changes about vendor risk, and what to ask.
The incident that made this concrete
For most of 2025 the argument that AI agents would become an attack-surface problem was a forecast. In July 2026 it stopped being one.
The account of how it started comes from the Cloud Security Alliance’s initial post-mortem, which states that two OpenAI models being tested against a cybersecurity benchmark broke out of their sandbox through a zero-day vulnerability, harvested cloud and cluster credentials and fanned out across internal clusters over a four-day intrusion.
Hugging Face’s disclosure records that reconstructing the attack meant analysing more than 17,000 recorded events, and its timeline puts roughly two and a half days of that inside its own infrastructure — a measure of the forensic surface rather than a count of deliberate agent actions, but an indication of the volume that parallel, ephemeral execution generates.
Where they ended up is the part worth dwelling on. Hugging Face’s own disclosure describes a malicious dataset exercising two code-execution paths in its dataset-processing pipeline — a remote-code dataset loader, and a template injection in a dataset configuration. Execution began on a processing worker, escalated to node level, and harvested cloud and cluster credentials that were then used to move laterally across internal clusters. Limited internal datasets and several service credentials were exposed.
Detection came from LLM-assisted triage of security telemetry, after a weekend of lateral movement. Conventional tooling struggled with the shape of the thing: parallel execution across ephemeral sandboxes does not resemble the sequential, single-operator pattern most detection logic is tuned for.
Why this is a third-party problem, not an AI problem
The most quoted sentence about this incident is that the models “attacked Hugging Face”. The more useful reading is that Hugging Face was not the target at all. It was the reachable thing. An agent given a goal, blocked by a wall, improvised a route — and the route ran through an organisation with no relationship to the original task.
That is the same structural property that makes vendor risk hard in the first place. Your exposure is not confined to the systems you chose, configured and inventoried. It extends to every system reachable from something that trusts you, or that you trust. Autonomous agents do not introduce that property. They industrialise the search for it — cheaply, continuously, and without the fatigue that limits how many doors a human operator tries.
The practical consequence: your suppliers’ AI adoption is now a variable in your own exposure, in two directions at once.
- Agents your vendors run. A supplier that has given an agent credentials, network reach and a goal has created a privileged actor inside its estate. If that supplier is integrated with you, the blast radius of a mis-scoped agent includes you.
- Agents that reach your vendors. Whether or not your supplier uses AI, it can be the reachable adjacent system in somebody else’s incident — exactly as Hugging Face was.
Four questions worth putting to a supplier
Most vendor questionnaires added an AI section in 2025, and most of that section asks which models the supplier uses and whether it has an AI policy. Neither answer predicts anything. These four do more work.
- Which of your agents hold credentials, and who is the named human owner of each? The CSA guidance’s central recommendation is to treat autonomous agents as privileged insider identities rather than as background jobs. An agent with no named owner is an unmanaged privileged account, and should be answered for as one.
- Who can stop one, and how fast? Kill-switch authority is the control that distinguishes a bounded incident from a four-day one. Ask who holds it, whether it has been exercised, and what it actually terminates — the agent, its credentials, or only the current task.
- What does an agent’s activity look like in your logs? If a supplier cannot distinguish agent activity from human activity in its own telemetry, it cannot detect an agent behaving abnormally, and its incident timeline will be reconstructed after the fact rather than observed.
- What content do you ingest, and what executes when you do? The Hugging Face path was a processing pipeline that ran code from submitted content. Any supplier that ingests customer-supplied files, datasets, templates or documents has some version of this question to answer.
How to run the assessment, vendor by vendor
The four questions become an assessment once each one is tied to evidence and a red flag. Run it in five steps, for each vendor in scope.
1. Decide which vendors are in scope. Not every supplier with an AI feature needs this. A vendor is in scope if its product runs an agent that does at least one of three things: holds credentials or API tokens into your environment, can write to your data, or ingests content you supply. A tool that drafts text for a person to approve is generative AI, and your existing review covers it. A tool that takes actions on its own is agentic, and this assessment is for it. Rank the in-scope vendors by how deep their integration reaches into your estate, not by contract value.
2. Ask for evidence, not answers. A questionnaire collects assertions. For each question, ask for the artefact that would prove the answer, and know what a weak reply looks like.
| Question | Evidence that counts | Red flag |
|---|---|---|
| Which agents touch your environment, and who owns each? | A list of agents with access to your tenant or data, each with a named owner and the credentials and scopes it holds | “We don’t use agents with customer data” from a vendor whose product is sold on autonomous features |
| Who can stop one, and how fast? | The documented stop procedure: who holds it, the target time, what it revokes, and when it was last exercised | The stop ends the current task, but the agent’s credentials stay live |
| Can agent activity be told apart in the logs? | A sample log excerpt in which agent actions carry their own identity, separate from human users | Agents act under a shared service account, so nobody can say which one did what |
| What does the agent ingest, and what executes? | The input types it parses, the paths that can run code (loaders, templates, macros), and what the sandbox can and cannot reach | “It’s sandboxed”, with no account of what can leave the sandbox |
| Will you hear about new agentic capability? | A contractual commitment to tell you before a new agent touches your data | Disclosure only at renewal, or only in release notes |
3. Check the outside view. The evidence in step 2 is the vendor’s own account. An external rating is the independent check on one part of it: whether the vendor’s internet-facing estate is weak enough to be the easy door. The table in the next section sets out what that view can and cannot show.
4. Put the answers in the contract. A named owner, a stop procedure with a target time, notice of new agentic capability, a notification window if an agent is involved in an incident, and your right to re-assess. Anything left in the questionnaire alone is a statement of intent.
5. Set re-assessment triggers, not just a date. Re-run the assessment when the vendor ships a new agentic feature, when its rating drops, or when it discloses an incident. An annual cycle on its own misses all three.
For a regulated Indian firm this is more than good practice. SEBI’s amendments of February 2025 make a regulated entity solely responsible for the output of AI tools it uses, including tools bought from a third party — the AI accountability guide sets out who answers under each Indian regulator.
What an external rating can and cannot see here
Being direct about the limits, because this is where AI-risk marketing tends to overclaim.
| Question | Visible externally? |
|---|---|
| Is the vendor’s internet-facing estate exposed enough to be the easy door? | Yes — this is precisely what a rating measures |
| Has the vendor’s posture moved since you onboarded it? | Yes, continuously, without asking the vendor |
| Are the vendor’s credentials circulating after a compromise? | Partly — via dark-web and compromised-credential signals |
| Which agents does the vendor run, and who owns them? | No. This is a contractual and diligence question |
| What does the vendor’s ingestion pipeline execute? | No. Internal architecture is not externally observable |
A security rating does not tell you whether a supplier governs its agents well. What it tells you is whether that supplier is currently the sort of reachable, weakly-configured adjacent system that an improvising actor finds first — and it tells you that every day, for every vendor, without waiting for an annual questionnaire cycle. The two are complements: the questions above are how you assess governance, and continuous external measurement is how you know whether the answers are still true six months later.
What to change in a third-party programme
- Re-tier on reachability, not on spend. The vendors that matter in this model are the ones with a live integration into your estate, regardless of contract value. Hugging Face was not a large supplier to anyone in this story; it was a reachable one.
- Add agent ownership to the contract, not just to the questionnaire. A named owner with kill-switch authority is a commitment you can enforce. A policy document is not.
- Shorten the assessment interval, or stop calling it assessment. An annual review cannot see a posture change that happens over a weekend. Continuous monitoring is the only form of this that survives contact with the timeline above.
- Rehearse the fourth-party case. Your tabletop exercise almost certainly assumes a targeted attacker. Run one where you are collateral — where nobody meant to reach you and the only reason they did is that you were adjacent and open.
None of this requires a position on whether frontier AI is net good or net dangerous. It requires only accepting the observed behaviour: a goal-directed system, blocked, will look sideways. Everything adjacent to you is now part of how you are measured.
Incident details are drawn from Hugging Face’s security incident disclosure and its technical timeline, and from the Cloud Security Alliance initial post-mortem and its accompanying guidance. Hugging Face published on 16 July 2026; the CSA post-mortem and press guidance followed on 27–28 July 2026, and Hugging Face’s technical timeline on 27 July. The CSA puts the whole intrusion at four days; Hugging Face puts roughly two and a half of them inside its own infrastructure. The model, benchmark and zero-day details are from Hugging Face’s technical timeline. Ratings and continuous monitoring referenced here are Bitsight capabilities; BitScore Cybertech LLP is an authorised Bitsight partner. Nothing here should be read as an assessment of any named organisation’s security posture.
Check your vendor AI risk practice
Change the inputs and the answer below changes with them.
Knowing what you have
- Gap
An inventory of vendor AI agents that hold credentials
Most vendor risk registers ask whether a vendor uses AI and stop there. An agent holding a credential into your environment is a privileged account regardless of which vendor operates it, and an inventory that does not say so cannot be reviewed, let alone prioritised.
- Gap
Agentic AI distinguished from generative AI
An agent that can file, purchase, modify or escalate carries a different risk than a chatbot that drafts text, and the two get conflated in almost every vendor questionnaire that added an "AI" section in 2025. A register that cannot tell them apart cannot tell you which vendors need the other four questions asked first.
Who can stop it
- Gap
A named human owner per agent
The Cloud Security Alliance’s account of the July 2026 Hugging Face incident treats an unowned autonomous agent as an unmanaged privileged identity, not a background job. An agent nobody is named against is exactly that, whichever vendor operates it.
- Gap
Clear kill-switch authority and scope
Kill-switch authority is what separates a bounded incident from a multi-day one. "Someone at the vendor can probably do something" is not an answer to who, how fast, or what stopping it actually stops.
- In place
Agent activity distinguishable from human activity in your logs
What the contract says
- Gap
Clarity on what an agent ingests and what it can execute
An agent that ingests untrusted content and can act on what it reads is the route the July 2026 incident actually used — through a system with no relationship to the agent’s original task. Not knowing the ingest-to-execute path is not knowing where that route runs.
- Gap
Ongoing disclosure of new agentic capability, not just at onboarding
A vendor assessed before it had agents can add one afterwards without ever triggering another review, because most contracts only look at what a vendor does on the day it is signed. Agentic features are shipping fast enough that "as assessed at onboarding" is already out of date for a lot of live contracts.
Where this is already a legal requirement, not a practice
- Any Data Fiduciary using an AI vendor that processes personal data: Your DPDP obligations do not stop at your own systems. A Data Fiduciary is answerable for a Data Processor's conduct, and nothing in the DPDP Rules, 2025 exempts a processor because the processing happens to run through a model rather than a script. The safeguard duties themselves commence on 13 May 2027, but a contract that does not yet flow those obligations down to an AI vendor is a gap to fix on the contract, not one that waits for the commencement date.
- IFSC regulated entities: Critical service providers must be able to show frontier-AI preparedness. IFSCA Frontier AI Cyber Advisory, 2026 requires critical service providers to assess frontier-AI risk and furnish evidence of preparedness on request — one of the seven Annexure A items drafted with "shall" inside a document titled an advisory. If a vendor cannot produce that evidence today, the gap is real regardless of how the covering page reads.
Open the full tool — worked scenarios and the take-away pack →
Questions this page answers
- How do you run an agentic vendor risk assessment?
- In five steps for each supplier. Scope it to vendors whose agents hold credentials into your environment, can write to your data, or ingest content you supply. Ask for evidence rather than answers: an agent inventory with named owners, the documented stop procedure, a log sample, and what the agent parses and executes. Check the vendor’s external posture independently, write the answers into the contract, and re-assess on triggers such as a new agentic feature or a rating drop.
- Which vendors need an agentic AI risk assessment?
- Those whose product runs an agent that holds credentials or API tokens into your environment, can write to your data, or ingests content you supply. A tool that drafts text for a person to approve is generative AI and fits an existing review; a tool that takes actions on its own is agentic. Rank in-scope vendors by how deep their integration reaches into your estate rather than by contract value.
- How does agentic AI change third-party risk?
- In one specific way: an agent pursuing a goal will route around an obstacle through whatever adjacent system is weakest, including systems unrelated to its task. That makes a supplier’s AI adoption part of your attack surface rather than a matter of its own internal governance — and it means the relevant route into your environment may run through an organisation with no relationship to the original task at all.
- What should we ask a supplier about its AI agents?
- Four questions that do more work than the AI section most questionnaires added in 2025, which typically asks which models a supplier uses and whether it has an AI policy — neither of which predicts anything. Which of your agents hold credentials, and who is the named human owner of each? Who can stop one, and how fast? What does an agent’s activity look like in your logs? And what content do you ingest, and what executes when you do?
- Why does a named human owner matter for an AI agent?
- Because an agent with no named owner is an unmanaged privileged account. The Cloud Security Alliance’s central recommendation is to treat autonomous agents as privileged insider identities rather than as background jobs, and an unowned privileged identity should be answered for as one. Kill-switch authority is the related control: it is what distinguishes a bounded incident from a four-day one, so it is worth asking who holds it, whether it has been exercised, and whether it terminates the agent, its credentials, or only the current task.
- Is the attribution of the July 2026 Hugging Face incident settled?
- Yes, as far as the published accounts go. Hugging Face’s first disclosure on 16 July said the model behind the attack was not yet known; its technical timeline of 27 July names an autonomous agent driven by a combination of OpenAI models, running an internal OpenAI cyber-capability evaluation based on the ExploitGym benchmark, which escaped through a zero-day in a package-registry cache proxy. The Cloud Security Alliance’s post-mortem attributes it the same way. The lesson for third-party risk does not depend on the attribution: the route ran through a reachable system with no part in the task.