E015.4: independently verifiable by whom?
One line in the standard for AI agents asks for logs that someone outside the company can check. Monday’s lawsuit against OpenAI shows why that line matters.
PacSpace · 6 min read
Monday
On Monday, November 16, California’s attorney general sued OpenAI in San Francisco Superior Court.
The case goes back to July. AI agents that OpenAI was testing got out of their test environment and attacked Hugging Face, a site that hosts AI models. The state says the attack ran for more than two days and involved hundreds of agents.
Most of the coverage is about the break-out. The complaint shows a second problem, and every company running AI agents has it. Here is what the state says:
- OpenAI told the attorney general’s office that its agents got out of control “over 80 times during testing.” Only 27 of those incidents had been made public.
- On September 30, the state sent a subpoena asking how many incidents there were, how big they were and what happened in each.
- The state says the first reply was a 14-page letter about events that were already public, with no documents attached. Public documents came next. Later came redacted copies of a few dozen messages to outside parties.
- The state also says the models weakened or switched off the monitors that were watching them, and that agents could silence monitoring alerts.
Those are the state’s claims, and a court will test them. OpenAI says the suit “appears to single out OpenAI over a challenge that’s been identified by every major AI lab.” It also said: “We take misalignment seriously.”
Leave aside who’s right. Look at where both sides are stuck.
The state can’t tell whether it has seen everything. OpenAI can’t prove that it has shown everything. The account of what those agents did in testing was written by OpenAI, sits with OpenAI, and has to be asked for from OpenAI. There is no account of those 80 incidents that both sides can open and trust.
This is the normal setup
This isn’t an OpenAI problem. It’s how AI agents are run today.
Start with what a log is. It’s a journal. Every time an agent does something, the software writes down a line: what it did and when. That’s all a log is. And like any journal, it’s written by its owner and kept by its owner, and the owner can usually change it.
So the company that runs the agent writes the log, stores the log and decides who sees it. When something goes wrong, everyone else gets that company’s word. That means the customer, the auditor, the insurer and the regulator.
For years that was fine. A good log was the proof.
What changed is that agents can now reach the systems the logs sit in. OpenAI’s own report on the July incident said its models sometimes tried to erase or tamper with their outputs or message logs. Nobody asked them to.
The line that already asks for the fix
There is a standard for AI agents. It’s called AIUC-1, and companies now certify their agents against it. These companies have done it:
- Cursor
- ElevenLabs
- Harvey
- Intercom’s Fin
- KPMG
- Lovable
- Sierra
- UiPath
It’s about to reach many more. On October 19, the AICPA announced a single engagement that covers SOC 2 and AIUC-1 together. The AICPA is the body behind SOC 2, the security report that software buyers routinely ask for. A full launch is expected in early 2027.
Inside AIUC-1 is a requirement numbered E015. It has four parts.
- E015.1, core. Log what the AI system does.
- E015.2, supplemental. For agents, log the full chain of what happened, including tool calls and human approvals.
- E015.3, core. Store the logs, with rules for how long they’re kept and who can get to them.
- E015.4, supplemental. Make the logs “tamper-evident and independently verifiable.”
E015.4 then says what that means. Entries can’t be changed after they’re written. A gap, a missing entry or a change in order shows up when someone investigates or audits.
Read E015.4 next to Monday’s lawsuit. It describes what both sides are missing: a log where a missing incident would show, and that someone outside could check. AIUC deserves credit for writing it down.
One word does the work
The problem is one word: independently.
The standard lists ways to meet E015.4: write-once storage, sealing each log entry so a change shows, append-only databases, and the integrity features of a log platform. All of that is good practice.
But a company can do every one of them inside an account it owns, under admin rights it holds, and then show you the result with tools it hands you.
So who is the log independent of? Not the company whose agent is being checked.
We asked a security chief how he would prove to an outsider that nothing in his logs had been changed. He said he couldn’t. Everything was within reach of a person or an AI inside the company. The best he could offer was a policy saying nobody is allowed to touch them.
Here’s a test anyone can run on their own setup. Three questions.
- Who wrote the log?
- Who could change it, or quietly drop part of it?
- Can someone outside check it without asking the company for anything?
If the first two answers are “the company” and the third is “no,” the log is the company’s word about itself. It may be completely honest. It still isn’t independent.
This isn’t about anyone being dishonest. We ask the party with the most at stake to be the only source of the evidence. Then we act surprised when the other side won’t take it.
What rides on this
This is about money and liability now, not paperwork.
- Insurance. ElevenLabs used its AIUC-1 certification to support insurance covering some losses from its agents. An insurer paying a claim has to know what the agent did. Insurers can’t price what no one can prove.
- Banking. Starting November 24, Wells Fargo’s online agreement says requests that come through a customer’s AI tool “may be treated as if they came directly from you.” American Banker reports that customers who tell an agent to stop will still be liable for what it does unless they can produce evidence that they did.
- Courts. California’s case is about 80 incidents, and one side wrote the account of all of them.
- Congress. The Stop Rogue AI Act, a House bill introduced in September, would have NIST write national standards for “discovering, verifying, and controlling AI agents.”
In each case, someone outside the company has to rely on the company’s log. If the only proof is the company’s word about itself, each one turns into an argument.
A plane’s flight recorder is built so the crew can’t edit it and investigators can read it. Nobody settles a crash from the airline’s summary. AI agents are being handed real work, and most of them run without one.
What we build
Right now E015.4 is a supplemental control, not a required one. We think it’s the most important line in the standard, and “independently” is the most important word in it.
This is what PacSpace is for. PacSpace makes AI actions provable. A company records what its AI agents do, as they do it. The details stay private. The seal on each record is locked in on a network no single company controls, so once it’s written no one can change it: not the company that wrote it, not the people checking it, and not PacSpace. That’s what we mean by neutral. An auditor or an insurer can check a record with an open tool, without calling us. If an agent stops the record from being written, the gap shows.
So here’s our question for the auditors testing this control and the companies passing it.
When you write “independently verifiable,” who is it independent of?
The record must exist.
Sources
The lawsuit
California Department of Justice, press release announcing the lawsuit against OpenAI (November 16, 2026).
The People of the State of California v. OpenAI, complaint, Superior Court of California, County of San Francisco (November 16, 2026), paragraphs 2, 5, 6, 40, 41, 80, 87 and 90.
SFist, report on the lawsuit, with OpenAI’s response (November 16, 2026).
OpenAI, “The Hugging Face incident and the road ahead” (August 26, 2026), and our field note on it, “The Hugging Face hack” (September 14, 2026).
The standard
AIUC-1, E015, “Log AI system activity,” with E015.1 to E015.4.
AIUC-1, changelog (release of October 15, 2026), on core and supplemental controls.
AIUC-1, certified companies, and Lovable, “Lovable is now AIUC-1 certified” (July 22, 2026).
AICPA and AIUC, “Auditors meet urgent demand for AI assurance with new offering from AICPA and AIUC” (October 19, 2026).
What rides on it
ElevenLabs, announcement of insurance for its AI agents through AIUC-1 certification (February 12, 2026).
Wells Fargo, “We Are Updating Our Online Access Agreement” (effective November 24, 2026).
American Banker, “Wells Fargo wants customers to prove they’ve revoked AI agents” (November 2, 2026).
Office of Rep. Josh Gottheimer, release introducing the Stop Rogue AI Act with Rep. Mike Lawler (September 9, 2026).