Autonomous AI exploit agents have crossed from theory into live attack traffic. Google Threat Intelligence Group found an exposed staging server running a multi-agent framework, called exploit_agent, that compared web application code before and after security updates, built working exploits and ran them with no human involvement — a first, the company says (The Register). Two days earlier, the FBI warned that attackers have weaponized an open-source framework called ARTEX since at least August 2026, and that some intrusions went from initial targeting to privileged access in under three hours (BleepingComputer).
That timeline is shorter than most managed security contracts assume. If your MSSP or MDR provider is still measured on a next-business-day response, the service level you are paying for describes a threat that no longer exists. This article sets out what to change in the contract, and what to ask before you renew.
Why the patch-to-exploit window collapsed
The old defender's advantage was time. A vulnerability was disclosed, you read the advisory, and you patched before anyone built a reliable exploit. Frontier models have removed that buffer. Google's researchers expect unsophisticated attacks needing little or no human interaction within three months, and sophisticated, fully autonomous campaigns within six (BleepingComputer).
McKinsey makes the same point from the buyer's side: the binding constraint has shifted from finding vulnerabilities to fixing them, and flaws can now be found at a pace no security or infrastructure team was built to absorb (McKinsey). That is a staffing and process problem as much as a tooling one. Autonomous AI exploit agents turn a disclosure into a deadline you cannot negotiate.
There is a volume problem too. Barracuda's Mythos Hype Index, cited by MSSP Alert, projected 77,000 CVEs in 2026, up from 49,000 in 2025 (MSSP Alert). Each one is a candidate for automated weaponization. No internal team triages that list by hand, which is exactly why buyers outsource it.
What autonomous AI exploit agents do to standard MSSP and MDR terms
Most managed security agreements were drafted around human attackers. Three assumptions break.
The first is dwell time. Contracts often imply that an intruder will sit quietly for days before doing damage, giving monitoring tools a window to catch them. When reconnaissance and lateral movement happen at machine speed, that quiet period disappears.
The second is attribution. Autonomous AI exploit agents do not map cleanly onto user, service account or application categories, so existing monitoring struggles to attribute what they do (MSSP Alert). If your provider cannot explain how it watches non-human identities in your environment, the contract is silent on a growing share of your risk.
The third is coverage. ARTEX clients integrate external large language models and turn attacker goals into commands executed by specialized agents, and the FBI says the framework is especially effective against overlooked systems such as partner portals, legacy services and third-party integrations (BleepingComputer). Ask whether those systems are in scope. Where autonomous AI exploit agents probe them, an exclusion is a gap you carry.
Seven clauses to add before you renew
Treat the renewal as a re-scoping exercise, not a price negotiation. Work through this list with procurement and legal in the room.
- Detection SLA in minutes, per severity. Define time from first observable event to a validated alert, separately for critical and high. Hours are no longer defensible for internet-facing systems.
- Notification clock, triggered by the provider's alert. Start the clock at detection, not at the end of triage, and require a named contact and a written record.
- Defined patch-to-exploit assumption. State the window the provider designs for, and require it to tell you when a threat exceeds it.
- AI-attack coverage in writing. Name the frameworks and behaviours in scope, and say what is out of scope.
- Non-human identity monitoring. Require discovery of agents, service accounts and API keys in your estate, plus alerting on privilege changes.
- Audit rights over the agent-action log. If your provider uses its own automation, you need a time-stamped record of what it did and when.
- Service credits tied to missed windows. Credits price small misses; keep termination rights in a separate clause.
On the last point, the guidance from autonomous pentest contract practice is blunt: a commitment needs an objective, buyer-verifiable trigger, or you should not pay for it (Stingrai). The same logic applies to detection promises. The agentic AI security contract terms that buyers are already drafting cover much of this ground.
Ask vendors to prove it, not describe it
The market is full of claims about AI-powered defense. Narrow them to evidence.
Ask for a walkthrough of one real incident where automated tooling found and contained an intrusion. Ask how the provider detects a burst of automated requests, unexpected credential changes and large transfers — the indicators the FBI recommends monitoring (BleepingComputer). Ask where detection rules come from, and how fast they are deployed.
A useful screen for vendors on detection quality is their testing regime. A provider that does not test its own detection content against new exploit frameworks is guessing. You can browse QA and testing specialists who work on exactly this kind of validation.
Red flags that should end a negotiation
Several signals tell you the provider has not adjusted its service to autonomous AI exploit agents.
- The response SLA is quoted in business hours or business days for critical alerts.
- The word "AI" appears throughout the proposal but no clause mentions AI-enabled attacks.
- The provider cannot show a log of its own automated actions in your environment.
- Non-human identities, partner portals and legacy systems are excluded from scope without explanation.
- There is no stated assumption for how quickly a published patch will be attacked.
If you are comparing proposals, the same discipline applies as in any vendor selection exercise: put the requirements in the RFP, then score answers against them. A vendor that answers with specifics is telling you something useful. A vendor that answers with brochures about autonomous AI exploit agents is telling you something too.
What providers need to change to keep the account
Providers face a harder sales conversation, and a genuine delivery one. The demand signal is there: 78% of UK and Ireland IT service providers plan to expand cybersecurity offerings over the next 12 months, and 47% name criminal use of generative AI as their top security concern, up from 37% in 2025 (ChannelPro).
The readiness gap is the risk. In OpenText's 2025 survey of over 1,000 MSPs, 92% said interest in AI was driving growth, but fewer than half felt fully confident guiding customers on autonomous agents (OpenText). Buyers now ask questions that expose that gap.
A practical response: publish your detection and notification commitments with per-severity times, describe how you watch non-human identities, and show a sample of the audit log clients receive. If you subcontract, say so and name the controls that flow down. The most credible proposals look like the agentic AI security contract terms buyers are already drafting against.
The same applies when a client wants agentic tooling on their side. Controls over AI agent permissions in their estate should be part of your managed scope, not a separate professional services line. Vendors who price the work as a fixed-price or time-and-materials engagement should show how detection and notification duties fit the model they propose. If you respond to tenders, the RFPs on the marketplace are where these questions now appear first.
Frequently asked questions
What are autonomous AI exploit agents?
They are AI systems that discover vulnerabilities, build working exploits and run them without a human approving each step. Google documented a multi-agent framework called exploit_agent doing exactly this against web applications in 2026.
How fast are AI-driven attacks now?
The FBI says ARTEX intrusions have gone from initial targeting to privileged access in under three hours. Google expects fully autonomous, sophisticated attacks within about six months.
Should an MDR contract include AI-specific SLAs?
Yes. Detection and notification times should be stated per severity, and the contract should say which AI-enabled attack behaviours are covered. Otherwise the provider can argue a machine-speed incident falls outside the agreed service.
What is a realistic patch-to-exploit window to plan for?
Treat it as hours, not weeks, for internet-facing software. Ask your provider what window it designs for, and require escalation when a live threat beats that assumption.
What this means for buyers
Renewal season is the moment to fix this, because the terms you sign now will govern the next twelve months of incidents.
- Pull your MDR or MSSP agreement and mark every response time quoted in hours or days.
- Replace them with per-severity detection and notification clocks measured from your notice.
- Add a clause naming AI-enabled attacks, non-human identities and overlooked systems in scope.
- Demand audit rights over any automated action taken in your environment.
- Ask three questions in the RFP: show the log, show the detection rule, show the last incident where automation helped.
Then test the answers against your own telemetry. If you cannot verify a promise, cut it and spend the money on a provider that offers one you can. The cybersecurity vendor directory is a sensible place to start a shortlist, and a posted project brief with these clauses attached will filter out most of the noise before the first call. As autonomous AI exploit agents keep shortening the gap between patch and attack, the contract is the part of your defense you can still fix on your own schedule.
- managed security contracts
- MSSP and MDR SLAs
- patch-to-exploit window
- AI attack coverage
- detection and notification terms
Sources
The pages this article draws on.
- 1.Google: AI agent turned a security patch into a working exploit with no human helptheregister.com
- 2.FBI warns ARTEX AI hacking tool is powering global cyberattacksbleepingcomputer.com
- 3.Google warns that fully-automated AI attacks are coming soonbleepingcomputer.com
- 4.How frontier AI is rewriting software security rules | McKinseymckinsey.com
- 5.Autonomous AI turns security risk into an MSP strategy issue | perspective | MSSP Alertmsspalert.com
- 6.Autonomous Pentest Contracts: Enforceable SLA Clausesstingrai.io
- 7.AI is rewriting the rules for MSPs in cybersecurity | ChannelProitpro.com
- 8.OpenText Cybersecurity 2025 Global Managed Security Survey: AI Redefines MSP Strategy | OpenText Blogsblogs.opentext.com
The Scopelyst team
Written by the people who run the marketplace, from what buyers and vendors do on it. No guest posts, no sponsored articles.
