AI GOVERNANCE
If your controls cannot tell a customer’s trusted agent from an intruder, your keenest adopters will be the first people you lock out.
The Write-Off: The intruder wore a company badge
THE WRITE-OFF · ISSUE 005
Meta’s security systems reportedly locked a small business out of Facebook after the owner handed his pages to Meta’s own AI agent. The agent then helped him sue Meta for $75,000. When your product strategy and your controls disagree about who counts as a customer, the customer pays first and the company pays later.
For several months, Joe Thomas, who owns Kong Koatings Epoxy Countertops in South Carolina, has let Meta’s Muse agent run parts of his business. He calls it “Baby Mia”, after his granddaughter, and it writes his emails and posts AI-generated videos to his company’s Facebook pages. According to 404 Media, once Thomas gave the agent access to those pages, Meta’s moderation systems disabled them because they suspected the agent was a hacker. Thomas markets entirely through Facebook ads, so for roughly a week his inbound leads dried up.
The agent did not sit quietly. It emailed journalists, contacted lawyers, filed complaints with the FTC and FCC, wrote to the FBI and the California Attorney General, set up a call with Senator Tim Scott’s office and, for good measure, faxed the White House. After the senator’s office referred the case, a Meta employee made contact and access was restored. On 16 November Thomas filed a lawsuit against Meta in Oconee County Court, which he says the agent drafted.
The complaint claims at least $25,000 in lost work and asks for $75,000, trebled under South Carolina’s Unfair Trade Practices Act, along with an explanation of $111.69 in ad charges. In fairness to Meta, the filing itself never names Muse as the cause; it describes a 7 November lockout behind a possible-hack warning, and password resets that kept failing after Thomas passed a selfie identity check. These are untested allegations, and no response from Meta appeared in the coverage.
THE VERDICT
The sums are trivial for Meta and painful for a countertop installer. A company pushing hard for agent adoption appears to have left its account-integrity controls tuned for a world where only humans log in. When the collision came, the recovery route ran through a senator’s office, and the customer’s agent turned a support ticket into a lawsuit and a pile of regulatory complaints within about a week.
Meta’s product side wants agents acting for customers around the clock. Its integrity side exists to stop anything that logs in and behaves unlike its owner. Each team was presumably doing its job well. Nobody appears to have owned the moment a sanctioned agent met a security rule, which is precisely where the customer got hurt.
An agent posting at speed, from new contexts, on someone else’s behalf, has much the same fingerprint as a compromised account. Unless the agent carries its own declared identity and scope, controls can only guess, and a cautious guess means a lockout. Borrowed credentials make delegation hard to tell apart from theft.
Thomas reportedly passed a selfie check and still could not reset his password. Restoration followed a political referral. The information-to-decision loop for a wrongly blocked customer was effectively missing, so the decision took a week and outside intervention. A small business does not have a week of leads to spare.
Customers now have agents that draft complaints, cite statutes and contact regulators without tiring. A grievance that once ended in a grumpy forum post now arrives as a filed lawsuit seeking treble damages. Support and legal teams sized for human-paced complaints will find both the volume and the formality have changed.
Thomas ran all his marketing through one platform and, by his account, much of his operations through an agent hosted by the same company. When that company’s control fired, both failed together. That concentration is his risk to manage, and it is also a mirror for any SaaS business whose customers rely on it end to end.
The principle is easy to state. If you encourage customers to delegate work to AI, your safeguards must know a delegate when they see one. Deterministic by default, AI by advantage, resilience by design applies to your controls as much as to your products.
Give every agent its own identity. Agents should act under declared, scoped credentials that your security, fraud and moderation systems can read, so delegation is visible and revocable instead of hiding behind a human login.
Make one executive own the collision. Whoever sponsors agent adoption should also sign off how every control treats agent traffic, with integrity teams consulted before launch.
Count false positives as a cost. Track lockouts of legitimate customers on a process behaviour chart, and treat a shift after an agent launch as a signal that needs a decision.
Design the way back in. Wrongly blocked customers need a human route to recovery with a time standard measured in hours. If the escalation path is a politician, there is no escalation path.
Expect disputes to arrive pre-drafted. Assume complainants have agents too, and size support and legal capacity for claims that are faster, more formal and more frequent.
The fixes are mostly organisational and cheap. Putting the people who sell agents and the people who stop intruders in the same room before launch costs far less than meeting later in Oconee County.
THE QUESTION FOR MONDAY’S LEADERSHIP MEETING
If nobody in the room knows, a customer’s agent may be drafting the answer already.
NEWTHISTLE NUGGET
AI GOVERNANCE
If your controls cannot tell a customer’s trusted agent from an intruder, your keenest adopters will be the first people you lock out.
The Write-Off: The intruder wore a company badge
SOURCES
A Man and His Muse AI Agent Sue Meta, 404 Media, 17 November 2026
Meta’s Own AI Agent Helps Businessman Sue Meta for $75,000 After Facebook Mistook It for a Hacker, IBTimes UK, 18 November 2026
A South Carolina countertop maker says his Meta AI agent fought to unlock his Facebook pages, then helped him sue Meta, Digital Wildcard, 18 November 2026
Meta’s own AI helps man sue its creators after it got him banned from Facebook, Dexerto, 17 November 2026