The Theft Looks Like Work
Economic espionage against American manufacturers does not arrive as an intrusion. It arrives as an employee doing something slightly unusual on a system he is allowed to use.
In January 2026 a federal jury in San Francisco convicted Linwei Ding, a former Google software engineer, on seven counts of stealing Google’s AI trade secrets. The mechanics are worth sitting with, because they are ordinary. Between approximately May 2022 and April 2023 he moved more than two thousand pages of confidential material into his personal cloud account — the architecture of Google’s custom AI chips, the design of the supercomputing data centers around them, the network hardware that ties them together — while involving himself with two technology companies in China and applying to a government talent program in Shanghai.
The same jury convicted him on seven counts of economic espionage. In August the trial judge threw those out. Judge Vince Chhabria found the evidence insufficient for a rational jury to conclude that Ding knew or intended his conduct would benefit the Chinese government. The theft convictions stood. He called the conduct a systematic, brazen effort to steal Google’s property, and in September sentenced him to just under a year in prison, two years of supervised release, more than $189,000 in restitution and a $25,000 fine.
That split is the problem in a single case. The taking was straightforward to prove, because it happened on a network that recorded it. The purpose was not, because nothing in the conduct announced it.
No firewall was breached. No malware was deployed. An employee with legitimate access moved files he was allowed to touch into a place he was allowed to use.
That is what this threat looks like from inside a company, and it is why the security controls most manufacturers buy do not catch it.
The approach arrives through the professional network
The recruitment side is equally unglamorous. A Five Eyes bulletin issued in June 2026, Safeguarding Our Secrets, describes China’s military intelligence services posting job advertisements on LinkedIn, Indeed and Upwork under the names of cover companies — private consultancies, think tanks, HR firms — and ranking the resumes that come back by how likely the applicant is to have access to sensitive information. The named targets are clearance holders in defense, foreign affairs and intelligence, military personnel, and people with peripheral access: academics, journalists, freelance writers, think tank staff.
Anyone who has worked counterintelligence recognizes what follows, and the bulletin lays it out plainly. The interview is virtual and the recruiter conceals who he actually works for. The first task is a trial report on something unclassified — China’s bilateral relations, the Indo-Pacific, defense or trade — and it pays a few hundred dollars. Then the client needs something more privileged, the conversation moves to an encrypted app, and the rate goes up. Payment often arrives from an account belonging to someone the recruit has never met. By the fourth report the relationship itself is the leverage, and the person on the other end knows exactly how much it would cost the target to explain the first three.
Two implications follow for a company. First, the approach happens off your network, on a platform your security team does not monitor and should not. Second, the moment where intervention is cheap is early — while the target still thinks of the contact as a recruiter or a client, and before a single file moves.
That window only exists if an employee has somewhere to report a strange approach without feeling they have put their own career at risk. Most companies have no such channel, and the ones that do rarely tell anyone it exists.
Not only China, and not only the United States
Two cases outside the US make the point that this is an industry condition rather than a single bilateral problem.
In July 2025 the Rotterdam District Court sentenced a former ASML engineer to three years in prison, convicting him of computer hacking and of violating EU sanctions by providing technical assistance to Russia. Between May 2023 and August 2024 he passed files taken from both ASML and NXP systems to a contact in Russia, using Signal, Telegram and Google Drive. Prosecutors could not prove he had sold them.
In December 2025 Taiwan’s High Prosecutor’s Office announced indictments against a subsidiary of Tokyo Electron and several associated individuals over the alleged theft of advanced TSMC process technology. Charges were brought under both the Trade Secrets Act and the National Security Act, the latter amended in 2022 to protect designated “national core critical technologies” as national security assets rather than ordinary intellectual property. Lawyers tracking the case describe it as the first significant corporate indictment under that framework.
Different countries, different adversaries, same target class: the process knowledge that makes advanced manufacturing possible. And note the second case carefully. The defendant is a supplier, not a state. In this industry your counterparties sit inside your process.
Why the tooling misses it
Most insider programs are bought as software and deployed against the wrong model of the problem.
Data loss prevention is tuned for volume and for prohibited destinations. Ding’s uploads went to a cloud account, over months, in quantities that individually looked like nothing. Access reviews ask whether someone is authorized. He was. Behavioral analytics flag deviation from baseline, but an engineer reading engineering documents is the baseline.
What actually distinguishes these cases is a change in behavior paired with a change in circumstance: unusual access alongside a new outside business interest, foreign travel that does not match the role, a resignation that follows a burst of downloads, a name that turns up on a foreign talent program application. No single one of those is evidence of anything. Together they are a pattern, and no tool assembles that pattern on its own, because the pieces live in HR, in legal, in travel records and in security systems that do not talk to each other.
What a real program requires
The federal model is not secret. The National Insider Threat Task Force, co-led by the National Counterintelligence and Security Center and the FBI, has published guidance for years, and its central design principle is the one companies most often get wrong: a program targets anomalous activity, not individuals.
In practice that means four things.
A standing multi-disciplinary panel. Security, HR, legal and, where relevant, counsel for privacy and employment law, meeting on a schedule to look at aggregated indicators. Not a tool. A group of people with the authority to ask the next question.
Protections built in from the start. The federal programs coordinate with records management, legal counsel and civil liberties and privacy officials specifically so that monitoring does not become surveillance of protected activity. A corporate program that skips this step will be shut down by its own general counsel the first time it produces an awkward result, and it will deserve to be.
A reporting channel for approaches. Separate from the ethics hotline, understood by employees to be non-punitive, and used by leadership to make clear that being contacted is not a mark against you but staying quiet about it is.
An exit process that assumes departure is the risk window. The download burst, the copied drive and the “personal archive” happen in the last ninety days, and often after the person has already accepted the next offer.
The point
An intrusion is a thing that happens to your network. This is a thing that happens to your people, on systems they are supposed to use, over months, in ways that look like work right up until the day they do not.
And the deterrent is thinner than a board might assume. Two thousand pages of frontier AI infrastructure design produced a sentence of under a year.
Companies building fabs in the United States are being asked to carry a class of risk that used to belong mostly to governments. The construction budget is public, the technology is strategic, and the adversary has already demonstrated what it is willing to spend to acquire it. The security organization that comes with that is built on the human side of the problem, and that is a discipline, not a purchase order.
Sources
- U.S. Department of Justice, “Former Google Engineer Found Guilty of Economic Espionage and Theft of Confidential AI Technology” (30 January 2026) — the January verdict and the underlying conduct: more than 2,000 pages uploaded to a personal cloud account between approximately May 2022 and April 2023. The release predates the August 2026 acquittal on the economic espionage counts and has not been amended; cite it for the conduct and the trade-secrets verdict only. justice.gov
- Reuters, “Ex-Google engineer’s conviction for stealing AI secrets partially overturned” (20 August 2026) — Judge Vince Chhabria, Northern District of California, vacated the economic espionage counts for insufficient evidence that Ding intended or knew his conduct would benefit the Chinese government; the theft-of-trade-secrets convictions stood. reuters via wsau.com
- KQED, “Ex-Google Engineer Gets Nearly 1 Year for Stealing AI Secrets” (1 September 2026) — just under a year in prison, two years of supervised release, more than $189,000 in restitution and a $25,000 fine; the judge’s “systematic, brazen effort to steal Google’s property.” kqed.org
- “Safeguarding Our Secrets,” Five Eyes bulletin (June 2026) — job ads on LinkedIn, Indeed and Upwork placed by cover companies; resumes ranked by likely access; virtual interviews with concealed identities; a paid trial report, then escalation to privileged material over encrypted messaging at a few hundred to several thousand dollars per report. The document names the FBI for reporting and refers throughout to Five Eyes agencies; it carries no printed date or reference number. ic3.gov
- The Register, “Ex-ASML engineer jailed for three years” (10 July 2025) — Rotterdam District Court; convicted of computer hacking and of violating EU sanctions by providing technical assistance to Russia; files from ASML and NXP shared via Signal, Telegram and Google Drive between May 2023 and August 2024; prosecutors could not prove the files were sold. theregister.com
- Pillsbury Winthrop Shaw Pittman, Global Trade & Sanctions Law, “Taiwan Prosecutors Announce First Major Semiconductor Indictments under New National Security Framework” (26 January 2026) — law firm analysis of indictments announced 2 December 2025 by Taiwan’s High Prosecutor’s Office against a Tokyo Electron subsidiary and several individuals, under the Trade Secrets Act and the National Security Act as amended in 2022. Not a primary government source. globaltradeandsanctionslaw.com
- National Counterintelligence and Security Center, “Insider Threat” (archived) — NCSC co-leads the National Insider Threat Task Force with the FBI; programs target anomalous activities, not individuals, and coordinate with records management, legal counsel and civil liberties and privacy officials. archive.dni.gov