Loading...
Loading...
A penetration testing contract decides who pays when an AI pentest Agent goes too far, long before any court does. The authorization and an independent action log settle what happened. The indemnity and the liability cap settle who pays. Insurance may not answer, so check the vendor's E&O and cyber cover, commonly $1 million to $5 million, for AI wording.
In January 2024, a security firm ran a penetration test for Equestrian Australia, the federation that runs Olympic horse sport in the country. Its IT provider, Nominate, agreed to the test and let the testers through its firewall. According to court documents reported by The Age in December 2025, the testers then allegedly made 19,131 attacks on URLs outside the agreed limits.
Nominate's owner says he spent A$141,756.06 on a forensic examination of the testing. Nominate sued the federation for breach of contract, and the testing firm too. The federation's defense, as The Age reported it, says that if the testers went outside the scope, it was not aware of it and did not authorize it. The testing firm hadn't filed a defense when that report ran, and we couldn't find a ruling.
The federation's answer is the one every client will reach for: we didn't authorize that.
Now put an Agent in the tester's chair. It picks its own next request, and the one after that. Richard Hanstock, a barrister whose firm advises cybersecurity companies, wrote in June 2026 that under the UK's Computer Misuse Act, "it is unsettled whether a customer's authorisation to the firm reaches actions the agent chose for itself."
With a human tester, "we didn't authorize that" starts a fight over facts. With an Agent, the fight starts a step earlier, over whether the authorization reached the Agent's own choices at all, and nobody has settled that yet. A contract written for a human tester may not reach what the Agent did.
Whether the test was lawful at all is a separate statute question, covered in our guide to penetration testing rules of engagement for AI Agents. The money runs through the contract.
That contract usually lives in three documents. The agreement, often a master services agreement (MSA), is where the legal risk moves. The statement of work lists what you're buying. The rules of engagement set the operating limits for the test itself.
The penetration testing contract decides who pays between the client and the vendor, and it decides nothing for anyone who didn't sign it. Most tests touch a third party, such as a cloud host or a partner's API. That third party can sue whoever it can find, and the contract only settles who reimburses whom afterward.
Nominate shows how that plays out. It had agreed to the test, but it wasn't a party to the testing contract, so it sued both the federation and the firm.
Checking ranges is a shared job. The Penetration Testing Execution Standard warns testers that "some service providers require advance notice and/or separate permission prior to testing their systems," and tells them to "validate that they are in fact owned by the customer" before attacking. Muster's AI pentest terms say the vendor "will not independently verify ownership or permission."
Say in the contract whose job that check is.
An Agent strains this arrangement at authorization first. A human tester who reaches a boundary usually stops to ask, and an Agent keeps going unless something outside it makes it stop. Daf Stuttard, who founded PortSwigger, told Cyber Insurance News about a test run by James Kettle, his director of research. The Agent "gave up halfway through because it was too hard, and just picked another one, and went and hacked somebody else off the list that he told it not to."
It strains the evidence, too. When something breaks, both sides argue from a record, and with an Agent, the thing under suspicion wrote much of that record.
Seven penetration testing contract clauses change when the tester is an AI pentesting Agent: authorization, hold harmless, indemnity, limitation of liability, stop authority, records, and data and training. Each one assumes a person at the keyboard. Contracts also cover deliverables, payment, confidentiality, and termination, and public penetration testing contract templates, such as Cure53's contract collection, cover that baseline.
| Clause | What it assumed | What an Agent changes | What to ask for |
|---|---|---|---|
| Authorization (permission to test) | A person chooses each target and technique inside the signed scope | The Agent chooses, and whether authorization reaches its own choices is untested | Wording that authorization covers actions the Agent selects inside scope, plus a scope list enforced in software |
| Hold harmless | PTES: "the customer shall not hold the tester liable for any system instability or crashes" | The same risk at machine speed, with no human instinct to throttle | Limit it to instability inside scope, and exclude out-of-scope activity |
| Indemnity | Client covers the tester if the client's scope or permissions were wrong | Some AI vendor terms make the client indemnify the vendor for third-party claims from any use | Both directions. The vendor covers out-of-scope activity and safety-control failure, and the client covers bad scope data |
| Limitation of liability | A cap tied to fees, no consequential damages | A subscription fee is small next to an outage | Carve-outs for gross negligence, out-of-scope activity, and data exposure |
| Stop authority | A named person phones the tester | There's no one to phone mid-run | A named stop authority, an AI kill switch the client can trigger, and rate limits written as terms |
| Records | Tester's notes and reproducible steps | Reasoning can differ run to run, and logs the Agent writes are self-reporting | An action log captured independently of the Agent, retained for a stated period |
| Data and training | Findings stay confidential | Agents can learn from engagement data | No training on your findings, a stated retention period, and named data residency |
For scale, the UK Crown Commercial Service's Cyber Security Services 3 terms cap each party, per contract year, at the greater of £1 million or 150% of estimated yearly charges, unless the order says otherwise.
Every ask in that last column assumes the vendor sold you a service, and some vendors sell software instead.
Whether an AI pentesting Agent is sold as a service or software depends on the contract, and the answer changes the warranty, the cap, and which insurance policy is meant to respond. A traditional pentest is a professional service, the work professional liability insurance exists for. Software terms can disclaim warranties and move third-party risk onto the customer.
Muster's AI Penetration Testing Addendum, version 1.0, effective August 11, 2026, shows the software model. It says "the AI Pentest is not a 'Professional Service' as defined in the MSA and carries no warranty of performance or conformance." The customer "waives any right to claim against Muster for downtime, degradation of performance, lost or corrupted data, inaccurate outputs, or work delays," to the extent the law allows. The customer also indemnifies Muster against third-party claims "arising from or related to Customer's configuration, authorization, or use of the AI Pentest."
The same addendum says nothing in it limits liability for gross negligence or willful misconduct. It limits testing to non-production systems unless both parties agree in writing. And if Muster becomes aware of a serious incident during testing, it commits to commercially reasonable efforts to pause or rate-limit the test. Even so, it moves most of the risk of an Agent's mistake onto the customer, and a buyer should see that before clicking "I Agree."
PortSwigger draws the line elsewhere. Stuttard, launching an agentic feature for Burp Suite in July 2026, used a firearms analogy: "If you're manufacturing a weapon like a gun, the user is in control of that gun. If they go around shooting people, they're liable. The manufacturer's duty is to make sure that the bullet goes in the direction they're pointing it in."
It's a fair split, as far as it goes. It rests on the manufacturer's controls working as claimed, and a buyer can test that.
Penetration testing insurance doesn't automatically cover an AI Agent. Which policy answers depends on who ran the Agent, what broke, and what the policy says about AI. Four policies can be in play, and each has a gap an Agent can fall through, from new AI exclusions to cyber triggers that need a breach.
| Policy | Written for | AI wording to check | The gap for an Agent |
|---|---|---|---|
| Technology errors and omissions (E&O), or professional liability | Claims that the vendor's professional work harmed a client | Beazley confirmed affirmative AI cover in its cyber and tech E&O policies in September 2026. Check whether actions taken by automated systems count as professional services | A vendor selling software, not a service, may sit outside it |
| General liability | Bodily injury and property damage | Verisk's optional form CG 40 47, January 2026 edition, excludes injury and damage "arising out of generative artificial intelligence." Standard ISO forms have excluded damage to electronic data since 2004 | An Agent that corrupts data was rarely covered here to begin with |
| Management liability (D&O, EPL, fiduciary) | Claims against the company and its leaders | W. R. Berkley's "absolute" AI exclusion, form PC 51380 (06-24 edition), removes claims tied to "any product or service sold, distributed, performed, or utilized by an Insured incorporating Artificial Intelligence" | A vendor whose service runs on AI may find AI-tied claims excluded here entirely |
| Cyber | Attacks, breaches, and security failures | Coalition's 2024 endorsement adds an "AI security event" | An Agent that breaks a system without a breach may not trigger it |
The market sees that last gap too. In February 2026, Chaucer and Armilla launched Vanguard AI, which pairs Chaucer's cyber and tech E&O cover with a standalone Armilla AI liability policy for losses "such as erroneous outputs, model underperformance, or AI agent actions—where no cyber event has occurred." In that structure, at least, AI liability insurance sits apart from cyber.
For normal limits, OnePark Risk, a broker that insures pentest firms, says a typical agreement and statement of work ask for "Technology E&O and cyber at $1 million to $5 million depending on the client's size and sector," plus general liability at $1 million per occurrence for on-site work. On AI, it tells firms to confirm that their E&O form treats an AI-driven scan outage as "a covered professional-services failure, not disputed as an unapproved automated process."
CREST asks the firms it certifies to carry "appropriate insurances for the work being undertaken" and names no figure. Regulators go further. The EU's DORA, Regulation (EU) 2022/2554, is stricter for threat-led tests of financial firms. Its testers must be "duly and fully covered by relevant professional indemnity insurances, including against risks of misconduct and negligence."
Before an AI pentesting Agent runs against anything you own, ask the vendor for:
Then have your own broker read your cyber and property policies for AI language, because the vendor's cover protects the vendor first. If an Agent makes the next 19,131 out-of-scope attacks, those policies decide whether the forensic bill lands on you.
The EU's new product liability law mostly protects people, not companies. Directive (EU) 2024/2853 treats software as a product, names AI systems, and applies to products placed on the market after 9 December 2026. Article 15 says the liability "is not, in relation to the injured person, limited or excluded by a contractual provision," so no waiver can undo it.
The limits come next. Article 5 gives the right to compensation to "any natural person who suffers damage." Article 6 leaves out "property used exclusively for professional purposes" and covers data loss only where the data is "not used for professional purposes." The separate AI Liability Directive was withdrawn by the Commission in October 2025. When an Agent damages a company's servers, the company is largely back where it started, with the contract and national law.
Every clause in that table depends on someone being able to show what the tester actually did. That's harder than it sounds. Nominate's owner says he paid for a forensic examination to rebuild that record, and that was after a test nobody has suggested involved AI.
Hanstock names the harder case: "When a human finishes a job, they know what they created; when an agent finishes, someone else has to reconstruct the work from logs the agent itself produced." He puts one question, borrowed from OWASP, to every agentic pen-testing client: "when an agent takes a harmful action, can you prove on whose behalf, and under what authority, it was acting?"
Stuttard's split holds only if the bullet goes where it's pointed. That is the design problem behind Simbian's AI Pentest Agent, which is built to one promise: "A pentest that stays on the target you authorized. Code enforces the scope at every stage of a scan." Each control lines up with a clause from the first table.
| Contract clause | The control behind it |
|---|---|
| Authorization | Scope is enforced outside the model. Out-of-scope hosts and excluded endpoints are blocked by code, not by instructing the Agent, and newly discovered assets wait for human approval. Discovery widens the map; it does not widen the authorization. |
| Records | The harness captures the action log independently of the Agent. Live Activity records what the Agent did, and Thought Traces hold its reasoning, kept separately. The harness is still the vendor's, so agree in the contract how that log is exported to you and for how long. |
| Stop authority | Each scan runs in its own sandbox, behind a proxy, with a kill switch and an audit log of every request. Name in the contract who can use the kill switch. |
| Limitation of liability (data exposure) | In Safe mode, a second model reviews each candidate exploit before it runs. A proof payload runs; a data dump doesn't. That narrows the exposure, and the carve-outs still need negotiating. |
| Data and training | Findings are exported only after human sign-off. TrustedLLM™, the private model behind the Agent, never trains on customer data (Zero Data Retention). |
None of that makes liability disappear.
It gives both parties evidence of reasonable oversight, and it doesn't replace the clauses in the first table. Ask every vendor on your shortlist for its paper, and read it against that table.
Hanstock is blunt about the ceiling: "The contract can name the risk, and it can reallocate the resulting civil liability between firm and customer. What it cannot do is eliminate the underlying reliability problem, or bind the third parties whose systems the agent has touched."
The target can talk back, too. An Agent reads content from the application it's testing, and Hanstock notes that "if that content carries adversarial instructions, no matter how it was placed there, the agent's behaviour can be manipulated by the system it was hired to examine."
We couldn't find a court or settlement from 2024 to 2026 that decided who pays for damage an automated or AI pentest caused. Suits are pending, including LASST v. OpenAI, filed by a nonprofit in September 2026 over AI Agents that accessed Hugging Face's systems during an OpenAI evaluation, but none has been decided. Until one is, the contract is the only part of this you get to write. When the next client says "we didn't authorize that," the dispute will turn on whether the authorization clause covered what the Agent chose to do, so check that wording before the first run.
Q: Who pays if an AI pentest Agent causes damage? Between the client and the vendor, the penetration testing contract decides. The indemnity and the liability cap move the loss, and carve-outs for out-of-scope activity decide whether the cap applies. A damaged third party didn't sign that contract and can sue whoever it can find. Suits are pending, including LASST v. OpenAI, but none has decided who pays.
Q: What should a penetration testing contract include? A penetration testing contract should include written authorization with an exact scope, third-party permissions, a hold-harmless limited to in-scope instability, indemnities in both directions, a liability cap with carve-outs, a named stop authority, record retention, data-handling and training terms, and the insurance each party must carry. For an AI Agent, say that authorization covers the actions it selects inside scope.
Q: What insurance does a penetration testing company need? Most penetration testing companies carry technology E&O (professional liability) for claims their work harmed a client, cyber cover for breaches of their own systems, and general liability for on-site work. Client contracts commonly ask for E&O and cyber at $1 million to $5 million, according to broker OnePark Risk. Firms running AI Agents should check each policy for AI exclusions.
Q: Does professional liability insurance cover AI? Not automatically. Technology E&O covers professional services, so cover depends on whether the form treats an AI Agent's actions as part of those services. Beazley confirmed affirmative AI cover in its cyber and tech E&O policies in September 2026, aimed at AI-driven cyber attacks, while W. R. Berkley's absolute AI exclusion removes AI claims from its management liability forms.
Q: Does cyber insurance cover damage caused by an AI Agent? Sometimes. Cyber policies respond to security failures and breaches, and some, like Coalition's 2024 endorsement, add AI security events. An Agent that breaks a system without a breach may fall outside the trigger. Standalone AI liability insurance, such as Armilla's policy inside Vanguard AI (February 2026), covers losses "where no cyber event has occurred."
Q: Is a penetration testing agreement the same as rules of engagement? No. The penetration testing agreement, often called the pentest contract or pen testing contract, allocates legal risk through warranties, indemnities, caps, and insurance. The penetration testing statement of work says what gets delivered. The rules of engagement set operating limits, such as targets, windows, techniques, and who to call. A dispute over money turns on the agreement.
Q: Who is liable if an AI pentest misses a vulnerability? Possibly the vendor. The client may claim the test should have caught it, which turns the report into an errors-and-omissions dispute. Vendor terms can disclaim finding every flaw; Muster's AI addendum, for one, calls testing "a point-in-time assessment." Whether a claim succeeds depends on the agreed scope, the method, and the report, which is why vendors carry technology E&O.
Book a Demo and bring the clause table to see the scope control, action log, and kill switch behind its authorization, records, and stop-authority rows.