Skip to content
Security & Trust

Prosecutors Say AT&T Hacker Sought AI Exploit Code While in Custody

Prosecutors say Wagenius used other inmates' email to ask a contact to prompt an AI tool for exploit code.

W
WebPulse Newsroom
AI-assisted · 3 min read
Share on X LinkedIn
Prosecutors Say AT&T Hacker Sought AI Exploit Code While in Custody

AI-generated image for WebPulse. About our images

Key finding

AT&T customers whose call and text metadata was stolen: 100 million+ (Source: KrebsOnSecurity (September 25, 2026))

A 70-month sentence, and a second set of allegations

Cameron John Wagenius, 22, was stationed with the U.S. Army in South Korea when he took on the alias Kiberphant0m. KrebsOnSecurity reported on September 25, 2026 that a Seattle court sentenced him to 70 months in prison and ordered $294,978 in restitution. The charges stem from 2024 intrusions at telecom carriers, including the theft of call and text records (numbers, timestamps, durations) for over 100 million AT&T customers. He had pleaded guilty on every count across two federal cases.

According to KrebsOnSecurity, the stolen data came from several large Snowflake customers that had leaked credentials and did not require multi-factor authentication on their accounts. The vendor now requires a second authentication factor on every account, Krebs notes. In October 2024, the persona claimed on cybercrime forums to have breached more than a dozen telecom operators, Verizon's Push-to-Talk business among them, and to be extorting them.

By the time alleged co-conspirator Conor Moucka was arrested, KrebsOnSecurity reports, AT&T had already handed the extortion group $370,000 in Bitcoin. The prosecutors' sentencing memo, filed September 19, puts Wagenius's own take from selling stolen data at around $1,500. The source does not say who received the ransom or how it was divided. Prosecutors conceded he was "not particularly financially successful" but argued that he intended to cause, and did cause, significant harm to individuals, U.S. companies and the government.

A note on how the case began, which is separate from the prison allegations below. Paul Russell, a resident agent in charge at the Defense Criminal Investigative Service, described the original investigation into the hacking and data trafficking. He told KrebsOnSecurity that his agency worked the case with the FBI, Army CID and the Secret Service after learning that a cleared soldier was allegedly tied to hacking tools and stolen data. Leads of that kind are rare, Russell said, and because investigators did not yet know the scope, they treated it as an "insider threat" from the outset. His comments concern that earlier phase, not the later Bureau of Prisons conduct.

What the memo says about the AI requests

The government's memo separately says Wagenius, held pending sentencing, broke Bureau of Prisons (BOP) computer-use policies by trying to learn about weaknesses in the agency's own systems. KrebsOnSecurity's summary of the same memo says he pleaded guilty almost immediately and has cooperated extensively.

According to BOP records cited in the memo, Wagenius used other inmates' email accounts to ask the email recipient to prompt a commercial AI tool. One request sought CVEs for Windows 10 Enterprise privilege escalation, with a working script for each. Less than a week later, from a different inmate's account, he asked for a step-by-step for CVE-2023-45208, a three-year-old command-injection flaw in D-Link networking devices, including code, or code written from scratch if none existed. Prosecutors also allege he asked for AI guidance on building an antenna from commissary items, and say he asked the recipient to research escaping prison.

In several instances, the memo says, the AI queries were framed as being in connection with a book he was writing. The government characterizes that as a common prompt-injection method: crafted inputs meant to get commercial AI tools, which are built to refuse to output code that exploits vulnerabilities, to comply anyway. The government told the court it is unaware of evidence that he worked out how to use or deploy any vulnerability he researched against BOP systems. When questioned, Wagenius said he was only researching "potential vulnerabilities to provide information to the BOP." The source does not say whether the recipients actually submitted the prompts, or whether any AI tool returned exploit code.

What this means for security budgets

This is one documented instance, alleged by prosecutors, of a person using a book-writing framing to try to obtain exploit code from a commercial AI tool. It is not evidence of a trend, and nothing in the source says the framing worked. The setting is a prison mail system, so the parallel to workplace AI assistants is loose. It still gives a concrete test case: a request for exploit code presented as fiction or research, possibly submitted by someone other than its author.

The original breach, by contrast, rested on a documented gap: customer accounts on a third-party data platform with exposed credentials and no enforced MFA.

Questions for your team

Ask security leadership whether MFA is enforced on every account connected to your cloud data platforms, including contractor, service and analytics accounts. Have the team test whether your AI-use policy and logging would flag a request for working exploit code framed as research or fiction, and whether they can see prompts submitted on someone else's behalf. Ask what happens when an insider is under investigation: which access is restricted, and how quickly could the team establish whether that person's activity ever touched production systems?

Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Krebs on Security.