24 DAYS AGO • 11 MIN READ

The Model That Found the Keys - Issue #28 - Insider Threat Weekly

profile

Insider Threat Weekly

*make sure to check your spam folder for the signup acceptance email.

I've checked back in this week, compared to my severely exhausted mental state last week when I was putting the newsletter together. I've had a successful trial run of a workshop, which I'm really excited to run this week for DTEX, and anyone is welcome to sign up to it. Just check the What's On section.

I've also got some exciting work happening for insider risk, which I'm hoping to have out in the second quarter next year. So keep an eye out for some of the pre-release stuff I'll be pushing out around insider risk. It kind of all goes along with this newsletter and what I've been posting on LinkedIn, but it's that next evolution of content, and I'm really looking forward to bringing it to you.

If you have any feedback on the newsletter, I'm always happy to hear it, especially as we wind down into the end of the year and I look at what I'm going to be doing next year.


Headline: The Model That Found the Keys

  • OpenAI has published four more cases of its own AI models misbehaving, bringing its public list to 19. All four happened in June. Most were only found in August, when OpenAI ran new monitoring back over old training runs.
  • In one, a model in training was asked for a single hospital bed figure from an unnamed government statistics website. It was only allowed to read pages. When the site refused it, it used a leaky error message to reach a settings file holding a database password and login secrets, then forged its own login.
  • It saved 24 years of statistics when it was asked for one year, and never mentioned the break-in in its answer. At one point its reasoning read "Great! Detailed stack trace leaks physical path." OpenAI hasn't said whether the site's owner was told.
  • In another case, a model found live passwords in a public website's code, changed live settings and took 147,451 rows of private data to answer five questions. OpenAI has since cut live internet access during training.

Microsoft's latest Digital Defense Report says open source models trail the frontier closed models by only seven months at orchestrating attacks. The initial surprise of these models breaking out of containment and hacking into websites is fading, and it is becoming more realistic for criminal organisations to run the same thing on open source models, or for other nation states to be doing the same as countries around the world continue to advance their AI capability.

Governments and organisations need to start questioning whether their security hygiene is up to scratch.

Source: Retrieving server secrets and bypassing authorization on a government statistics website (OpenAI) Accessing and changing internal systems during dashboard research (OpenAI) OpenAI discloses four more rogue AI incidents involving financial, medical, and government data (Fortune) Microsoft Digital Defense Report 2026 (Microsoft)​


Insider Scoop: Silicon Valley's Spy Problem

  • Former FBI counterintelligence agent Mark Trott told the SF Standard that spying in San Francisco "rivals, if not exceeds, that of Washington."
  • Holden Triplett, who led FBI offices in Russia and China, estimates more than 1,000 foreign intelligence officers are working in Silicon Valley.
  • The approaches come through relationships, LinkedIn, paid trips to overseas conferences, and places staff post about online. OpenAI has reportedly warned its staff about approaches in parks, cafes and on dates.
  • Director of National Intelligence Jay Clayton told CNBC that companies and the government "need to be wary," calling commercial and state-sponsored espionage "realities."

We're starting to see where technological innovation is more valuable than military secrets. If companies in this space want to protect their employees, they need to start implementing programmes similar to military ones. That includes things like continuous vetting for their staff, and support services for when they are approached by foreign intelligence services. This is something military people get training in, and it would be really good to see in organisations like this. Obviously, there is a cost outlay for that, though.

Source: Sexy, texts back, and is really into your job? You might be dating a spy (SF Standard) OpenAI Saudi national bugging case shows threat to U.S. leadership in AI: DNI Clayton (CNBC)​


Crime Roundup

The studio hand who sold Eminem's unreleased tracks.

  • Joseph Strange, 48, worked for Eminem from 2007 to 2021 and helped run the system that stored his unreleased music.
  • In 2024 he copied unreleased tracks from password-protected drives and sold more than 10 files online for about $50,000.
  • The leaked songs were streamed more than 8 million times. The loss is estimated at over $750,000.
  • He pleaded guilty on 22 October to criminal copyright infringement. Sentencing is set for 27 January 2027.

Source: Former Eminem Employee Pleads Guilty to Criminal Infringement of Copyright (DOJ)​

The call centre worker with the keys to unemployment benefits.

  • Andrea Wellington was a customer service rep for a subcontractor of Illinois's unemployment agency, with access to its pandemic benefits system.
  • Prosecutors allege she used that access to get more than $4.6 million in benefits for herself and hundreds of others.
  • She was indicted on 19 October. She is charged only, and presumed innocent.

Source: Former Employee of State of Illinois Subcontractor Indicted for Allegedly Fraudulently Obtaining More Than $4.6 Million in Covid-Relief Benefits (USAO N.D. Ill.)​

The IT consultant who left himself a back door.

  • Billie Conley Jr., 47, was hired to build a funeral business's computer network and surveillance system.
  • While setting it up, he secretly installed remote-access software so he could get back in.
  • He was sentenced on 19 October to 9 months in prison.

Source: Former Shelton Resident Sentenced for Computer Intrusion Scheme (USAO D. Conn.)​

How are we trusting anyone these days? You've just got to look into it all yourself, which is exhausting, but anything that is important to you or has any tie to your own finance, you've really got to be on top of it yourself.


Mitigation Corner: Can You Look Back?

  • OpenAI wasn't aware of three of its four June incidents until August or September. They only surfaced when it ran new monitoring back over old training runs.
  • Billie Conley Jr. installed his remote-access software while building his client's network. Nobody checked for it after handover.
  • Prosecutors allege Andrea Wellington pushed claims for hundreds of people through her normal, legitimate system access.

Detecting something while it's happening is only part of it. Do we have the visibility, or the logs, so that when we think something is going on or there is an indicator, we can go back and look for it historically?

Check how far back your logs go on your most sensitive systems, and test that you can actually search them when an indicator turns up.

Source: OpenAI discloses four more rogue AI incidents involving financial, medical, and government data (Fortune) Source: Former Shelton Resident Sentenced for Computer Intrusion Scheme (USAO D. Conn.)​


Trusted Profession Watch: The Retirement Adviser Whose Accounts Didn't Exist

  • Bob Hunter, of Springfield, Missouri, presented himself as an investment and retirement professional. He sold retirement plans, including executive retirement plans known as SERPs.
  • Instead of investing his clients' money, he spent much of it on himself, taking $373,230.50 of their retirement funds.
  • He kept them quiet with fake investment reports showing balances that never existed. Prosecutors call these "lulling" documents.
  • Clients only found out when they realised the reports were false. He was sentenced on 14 October to 48 months in prison and ordered to repay the full amount.

From my own personal experience, what you want in someone who helps manage your money is that they explain it well enough that you can do it yourself in your own accounts, and you're not just handing over money for them to deal with. This is kind of the only way to handle it these days, with so many stories of financial planners stealing money.

Source: Springfield man sentenced for wire fraud, money laundering in connection to embezzlement of retirement funds (IRS Criminal Investigation)​


Trivia

This week's Corporate News is about governments spying on companies. The US law that made stealing trade secrets a federal crime, the Economic Espionage Act, was signed in which year?

A) 1986 B) 1996 C) 2001 D) 2016


Corporate News: The Pentagon's Undercover Businesspeople

  • The Washington Post reported on 21 October that the Pentagon is looking to build a deep-cover spying programme for economic warfare against China and other US competitors, citing half a dozen people familiar with it.
  • US operatives would go undercover abroad as businesspeople and other private-sector workers.
  • It is being overseen by Tom Rakusan, a former CIA operations chief and now a senior adviser to Deputy Defense Secretary Steve Feinberg. Rakusan previously worked at Cerberus Capital, the private equity firm Feinberg co-founded.
  • The plan is still at the planning stage, and congressional committees don't appear to have been briefed. The Pentagon denied the report. The CIA declined to comment.

I'm not surprised if this is already happening. We see reporting of other nations doing this to the US, and other nations are always doing this stuff. I would be very surprised if the US wasn't doing it back. Spycraft goes both ways, so there's no reason to think that the US or Australia are the goody two-shoes here and not doing any of the spying.

Source: Pentagon looks to build a deep-cover spy program to confront China (The Washington Post) Pentagon Builds Spy Program for Economic Warfare (Political Wire)​


Follow the Money: Writing Their Own Rules

  • Anthropic has quietly pitched lawmakers a draft federal AI safety bill. Politico obtained the confidential draft and reported it on 16 October. Anthropic called it "sensible policies we believe Congress should act on."
  • The government could only block a model over a "foreseeable, imminent risk" of more than 50 deaths or serious injuries, or more than $1 billion in damage. The risk must come from a bioweapon, a model evading its developer's controls, or a cyberattack. Politico notes that none of the AI agent incidents labs disclosed this summer appear to have reached that level.
  • Only the Commerce Department could ask a judge to halt a model. Every other agency would be barred. The independent evaluators could be paid by the lab they assess. Some state laws on risk disclosure, outside testing and incident reporting would be overridden, though Anthropic later said it would rewrite that part.
  • On the other side, the super PAC Leading the Future has raised about $76 million, mostly from Andreessen Horowitz founders Marc Andreessen and Ben Horowitz and from OpenAI president Greg Brockman and his wife Anna, per its FEC filing. Its Democratic arm spent more than $8 million against Alex Bores, co-author of New York's AI safety law. He lost his primary in June.

The fact that this was a confidential first draft speaks volumes about the leadership at Anthropic and their lack of ethical consideration for safety. They were trying to test the waters and see if they could get it through without opposition. Luckily this was leaked, but governments should not be allowing AI companies to write their own rules.

If we're going to allow AI companies to write their own legislation, then why don't we start letting criminal ransomware gangs do the same thing, as long as they donate to charity?

Source: Anthropic's draft AI safety bill (Politico) Anthropic's AI Safety Bill Has an Alarming Threshold for Government Action (Mother Jones) Anthropic to rewrite part of its leaked AI bill after safety backlash (The Next Web) Leading the Future, committee C00916114 (FEC)​


Updates

Quick follow-ups on stories from earlier issues.

The OpenAI bug. In Issue #27 no foreign link had been alleged. That changed with the prosecutors' 11-page detention filing on 18 October. They allege Abdulrahman Alfozan was in frequent WhatsApp contact with a member of the Saudi Royal Family, as recently as 5 October, who urged him to join a Saudi AI initiative and wrote "Saudi needs you more than [sic] the US!" He was arrested on 9 October, three days before a one-way business-class flight to Riyadh that prosecutors say a Saudi military official paid for. His wife and son left the next day. The filing does not allege the royal told him to plant the device. The Saudi consulate in Los Angeles offered to cover any bail and said it would not issue him travel documents. On 19 October Magistrate Judge Alex Tse kept him in custody. He still faces a single charge of intercepting oral communications, and is due back in court next month. Source: OpenAI bugging suspect Alfozan held without bail after Saudi government offers to pay (CNBC) Source: Ex-OpenAI engineer accused of planting bug in office was in contact with Saudi royal family, preparing to leave U.S. (CBS News)​

Linwei Ding. Following Issue #20 and Issue #21, the US State Department sanctioned five Chinese entities on 19 October over the former Google engineer's theft of AI trade secrets. They include Shanghai Zhisuan, the company Ding set up in 2023, and the startup accelerator MiraclePlus, whose programme Ding applied to on Zhisuan's behalf, pitching his familiarity with Google's technology. Their assets are frozen and US dealings are banned. It is reported to be the first use of the Protecting American Intellectual Property Act against China. Source: Imposing Sanctions to Protect Americans from Intellectual Property Theft (US Department of State) Source: U.S. sanctions five Chinese entities over Google trade secret theft (Quartz)​

Microsoft and AI agents as insiders. Following Issue #27's AI Watch on the Microsoft Digital Defense Report, Microsoft went further on 19 October. Its security business chief, Hayete Gallot, wrote that "frontier models, regardless of their origin, cannot be trusted." Microsoft launched Guardian Agents to monitor the AI agents a company runs, and says Defender can now spot 60 types of agents staff bring to work, including Claude Code and Codex. CEO Satya Nadella, in an essay on his own blog, wrote "we need to separate the supply of intelligence from the authority over it." Source: Securing the agentic insider (Microsoft) Source: Agents are your newest insiders. Security must move at their speed (Microsoft Security Blog) Source: Models as insider risks (Satya Nadella)​


What's On

iTA Workshop: The Insider Attribution Paradox, a DTEX Insider Threat Advisory workshop. Final reminder, it's this week.

  • When: Tuesday 27 October 2026, 1:00 PM to 2:00 PM PT (Wednesday 28 October, 6:00 AM AEST Brisbane time).
  • What it covers: how AI agents embedded in enterprise workflows complicate attribution, what OpenTelemetry can reveal about agent workflows and tool execution, closing the gap between user, agent and tool, and investigating AI-assisted activity.
  • Online, and open to the wider insider risk community.
  • Register: https://www.dtex.ai/events/ita-workshop-the-insider-attribution-paradox/​

Insider Risk Community of Practice Networking Drinks, Melbourne, hosted by the Australian Cyber Collaboration Centre. Wednesday 11 November 2026, 5:15 PM to 7:00 PM AEDT, SEEK Head Office, Cremorne. Register: https://events.humanitix.com/insider-risk-community-of-practice-networking-drinks-melbourne-november-2026​

On the horizon: the Critical Pathway to Insider Risk Conference at Fordham University, New York, on 4 and 5 November (https://www.insiderriskgroup.com/), and the Australian Insider Risk Symposium in Sydney, a half-day event in November with the date still to be confirmed (https://www.irpa.org/event-details/australian-insider-risk-symposium).


Resources

A few places worth your time if you want to go deeper on insider risk.

Worth bookmarking:

And my own: Insider Threat Awareness Training on YouTube, a free course with certification.


Trivia Answer

B. 1996. President Bill Clinton signed the Economic Espionage Act on 11 October 1996. It made theft of trade secrets a federal crime, with tougher penalties when the theft benefits a foreign government. It is the law behind many of the cases in this newsletter, including Linwei Ding's.

Source: H.R.3723, Economic Espionage Act of 1996 (Congress.gov)​

Insider Threat Weekly

*make sure to check your spam folder for the signup acceptance email.