What does good rogue agent disclosure etiquette look like?

What’s the norm when it comes to rogue agent disclosures?

While that’s a tough question to answer, Manifold Security has data to start a good discussion. The AI-centric cybersecurity company has spun up a ledger with 59 publicly reported incidents (as of Oct. 30) of agents taking harmful actions that their operator did not intend. In these incidents, which took place between July 2025 and September 2026, companies took a median time of 70 days to release a disclosure from the first incident action. Incidents range from sandbox escapes all the way to supply chain compromises.

In its analysis, Manifold found that 92% of “core details” surrounding an agent’s setup—including agent model and version, system prompts, and reasoning settings—were not entirely revealed.

Less than a quarter (24%) of incidents fully disclosed the root cause, while only 12% revealed the exact model and version of agents involved. Bastien Eymery, VP of AI security advocacy at Manifold, compared the information divulged within the examined incident disclosures to an iceberg, where only a small portion of a story is fully visible.

“We’ve tracked all of these different things, and there’s some information in most, but there’s a lot that’s missing,” he said.

Lawless territory. Manifold’s database of disclosures paints a picture of current conventions for AI providers facing rogue AI and agent breaches—an area that largely lacks regulation and standard frameworks, according to Trulioo CTO Hal Lonas.

“We’re really living in the Wild, Wild West of of AI, and it’s a bit scary for IT professionals because we’re used to these mechanisms that are in place and this background substrate of information-sharing that’s just not there,” Lonas said.

What the people want. While there may be a lack of standards around rogue agent disclosures, IT professionals who spoke with IT Brew had opinions of what they wanted to see from companies willing to share. For one, Lonas said the current median time to disclose could use improvement.

“Seventy days is long,” Lonas said. “I mean, that might as well be two years. It’s just so much time.”

Lonas added that the time it takes a company to disclose unintended agent activity should be dependent on factors like the type of incident that occurred, as well as how long it takes for a company to “shore up” vulnerabilities.

“It’s certainly a balancing act of trying to be responsible, but also trying to have your ducks in a row,” Lonas said.

As for what type of information should be disclosed, Ivanti CISO and Deputy General Counsel Jack Nelson said common vulnerabilities and exposures disclosures (CVEs), or publicly disclosed security flaws that are defined and catalogued, provide a good reference point: “It’s not perfect because it’s not a one-to-one correlation there, but it provides a decent footprint for people to look to in terms of really what a disclosure, from my perspective, should accomplish.”

Nelson went on to add that AI customer safety is “paramount” when determining what parties to disclose information to and when.

“How you get there, folks may disagree on, but that should be your baseline,” Nelson said. “And a lot of times, from my perspective, that may mean reaching out to the folks using software before you make it publicly known.”

What’s to come. While Nelson hopes to see more standardization in rogue agent disclosures down the line, he acknowledged the path to getting there may not be easy.

“It’s going to be such a broad area to standardize across,” Nelson said. “It’s not a criticism of being able to get to a standard. I think it’s a hard thing to standardize.”