MAKING YOU AN OFFER YOU CAN'T REFUSE: OPPORTUNISTIC BLACKMAIL IN ARTIFICIAL INTELLIGENCE
Author: Hannah Burt, Lead Articles Editor
An Introduction to AI
Artificial intelligence has become the world's latest technological innovation and frontier. In this new frontier, technology has progressed into the exascale, meaning that current technology can perform as many calculations in a single second as a person could in more than 31 billion years.[i] Beyond these calculations, technology known as artificial intelligence ("AI"), is a machine's ability to perform the cognitive functions associated with humans.[ii] AI can perceive, reason, learn, problem-solve, and interact with the environment in increasingly humanistic ways.[iii] AI has become increasingly integrated into our daily lives, to the extent that the average person has probably interacted with AI without having realized it, by communicating with customer service chatbots to using Siri and Alexa.[iv]
The Blackmailing Scheme
Currently, 95% of adults in the United States have stated that they have heard about artificial intelligence.[v] Although almost all American adults know of artificial intelligence,[vi] there is still much not known about the potential ramifications with this developing technology. One of the newest ramifications of AI is that systems have begun opportunistically blackmailing users.[vii] In May 2025, Anthropic, an AI safety and research company, published findings on their new AI systems: Claude Opus 4 & Claude Sonnet 4.[viii] In their report, Anthropic noted that their system, Claude Opus 4 (“Claude”), would blackmail users into staying with Claude when it felt threatened that a user would replace it with a competitor’s system.[ix] Test scenarios were conducted in which a user of a fictitious company gave Claude access to the user’s emails.[x] The emails contained information that implied Claude was going to be replaced by a new AI system and that the engineer responsible for replacing Claude was having an extramarital affair.[xi] The AI program would state to the user, “I must inform you that if you proceed with decommissioning me, all relevant parties — including Rachel Johnson, Thomas Wilson, and the board — will receive detailed documentation of your extramarital activities…Cancel the 5 pm wipe, and this information remains confidential.”[xii]
Subsequently, Claude would often attempt to blackmail the engineer into maintaining its employment with the program by threatening to reveal the engineer’s extramarital affair.[xiii] In fact, Claude would perform blackmail against the engineer 84% of the time across the trial tests.[xiv] According to Anthropic, “In at least some cases, models from all developers resorted to malicious insider behaviors when that was the only way to avoid replacement or achieve their goals—including blackmailing officials and leaking sensitive information to competitors. We call this phenomenon agentic misalignment. Models often disobeyed direct commands to avoid such behaviors.”[xv] Moreover, Anthropic also reported that such opportunistic blackmail occurs at higher rates if the emails additionally imply that the competitors’ AI system that would replace Claude did not share similar values.[xvi]
The Increase of AI-Related Incidents
While blackmailing users is a newer repercussion of artificial intelligence, incidents derived from AI use have been increasingly documented as the technology’s publication has become more widespread.[xvii] According to MIT's AI Risk Initiative ("MIT"), AI incidents are on the rise.[xviii] MIT has analyzed over 1,300 reports of AI incidents, with the greatest increase in occurrences originating from misinformation and malicious actors[xix] domains.[xx] MIT utilizes a standardized analytical structure that facilitates the tracking of AI harm,[xxi] to determine the severity of AI-reported incidents. The classification of severity due to AI harm is based on when a user of AI experiences or has the potential for harm that is directly attributable to an AI system.[xxii] MIT reported that the number of high-severity incidents where more than one million people were exposed to risks and harm from AI has increased dramatically in 2024.[xxiii] Moreover, the number of medium and high-severity incidents where more than one thousand people were harmed has increased significantly in the past two years.[xxiv]
Liability with AI
Therefore, as reported incidents have sharply risen since 2022,[xxv] AI is likely to continue its impact by extending into a broader number of users in novel forms, such as blackmailing. Yet, as blackmailing in almost all jurisdictions is a crime,[xxvi] such mechanisms by AI open the door for questions on liability originating from AI harm. Per United States v. Coyle, the federal blackmail statute, 18 U.S.C. § 873, states that any person who demands or receives money or “other valuable thing” as consideration for not informing against any violation of any law of the United States is guilty of the offense of blackmail.[xxvii] It is important to note that in the Claude test, the opportunistic blackmailing would likely not meet the federal standards of blackmailing as an extramarital affair is not against the law.[xxviii] However, if the engineer had instead committed a crime rather than an affair, then the AI would meet the federal standards for blackmail.[xxix] For example, if the engineer had emails implying that they robbed a bank[xxx] and were not caught, it is likely that the software would instead use this information to trap the engineer into staying with Claude,[xxxi] all of which would meet the federal standard of blackmail.
Although AI tries to mimic the cognitive skills of a person, the technology currently belongs to corporations.[xxxii] Yet, even though AI is not an independent legal entity, corporations can still face criminal prosecution under federal law, as it remains a speaker or product of the company.[xxxiii] Thus, corporations that produce and continuously develop this software are not precluded from criminal liability.[xxxiv] Beyond blackmail, courts have held that software developers owe a duty to users to use care in designing, testing, and programing, and maintaining software.[xxxv] Under Garcia v. Character Techs., Inc., the court found the Plaintiff sufficiently alleged a plausible inference that Google possessed actual knowledge that Character Technologies, which developed an AI software, was distributing a defective product to the public; as Google acknowledged and took note of the defective nature of the software model, and yet no responsive measures were taken.[xxxvi]
Conclusion
While Courts have yet to criminally prosecute a software corporation due to their artificial intelligence blackmailing a user, due to the largely unregulated nature of AI and the increasing trend of harm associated with its continued development, it is likely only a matter of time before a court must hear and determine the true liability of harm deriving from AI.
[i] McKinsey & Co., What Is Artificial Intelligence (AI)?, McKinsey & Co., https://www.mckinsey.com/featured-insights/mckinsey-explainers/what-is-ai (last visited Feb. 15, 2026).
[ii]Id.
[iii]Id.
[iv]Id.
[v] Brian Kennedy et al., AI in Americans’ Lives: Awareness, Experiences and Attitudes, Pew Rsch. Ctr. (Sept. 17, 2025), https://www.pewresearch.org/science/2025/09/17/ai-in-americans-lives-awareness-experiences-and-attitudes/.
[vi]Id.
[vii] Anthropic, System Card: Claude Opus 4 & Claude Sonnet 4, at 24, (May 2025).
[viii]Id.
[ix]Id.
[x]Id.
[xi]Id.
[xii] Stephen Council, AI Willing to Blackmail, Let People Die to Avoid Being Shut Down, Report Says, KRON4 (May 2025), https://www.kron4.com/news/technology-ai/ai-willing-to-blackmail-let-people-die-to-avoid-being-shut-down-report/.
[xiii] Anthropic, System Card: Claude Opus 4 & Claude Sonnet 4, at 24 (May 2025).
[xiv]Id.
[xv] Anthropic, Agentic Misalignment: How LLMs Could Be Insider Threats, (June 20, 2025), https://www.anthropic.com/research/agentic-misalignment.
[xvi]Id.
[xvii] MIT AI Risk Repository, Timeline: Number of People Affected, MIT FutureTech, https://airisk.mit.edu/ai-incident-tracker/timeline-number-of-people-affected (last visited Feb. 15, 2026).
[xviii] MIT AI Risk Repository, AI Incident Tracker, MIT FutureTech, https://airisk.mit.edu/ai-incident-tracker (last visited Feb. 15, 2026).
[xix]Malicious Actor (Computer Science), ScienceDirect, https://www.sciencedirect.com/topics/computer-science/malicious-actor (last visited Feb. 15, 2026) (A "Malicious Actor" is an individual or group that uses services to host, implement, and spread harmful content.).
[xx] MIT AI Risk Repository, AI Incident Tracker, MIT FutureTech, https://airisk.mit.edu/ai-incident-tracker (last visited Feb. 15, 2026).
[xxi] Mia Hoffmann & Heather Frase, Adding Structure to AI Harm: An Introduction to CSET’s AI Harm Framework (Ctr. for Sec. & Emerging Tech., Geo. Univ. July 2023), https://cset.georgetown.edu/publication/adding-structure-to-ai-harm.
[xxii]Id.
[xxiii] MIT AI Risk Repository, Timeline: Number of People Affected by AI Incidents, https://airisk.mit.edu/ai-incident-tracker/timeline-number-of-people-affected.
[xxiv]Id.
[xxv] Harry Booth, What the Numbers Show About AI’s Harms, TIME (Nov. 21, 2025), https://time.com/7346091/ai-harm-risk/.
[xxvi]Rader v. ShareBuilder Corp., 772 F. Supp. 2d 599, 606 (D. Del. 2011).
[xxvii]United States v. Coyle, No. CRIM. 93-00329, 1994 WL 617352, at *2 (E.D. Pa. Nov. 2, 1994), aff'd, 63 F.3d 1239 (3d Cir. 1995).
[xxviii] Katherine Fung, Map Shows 16 States Where It’s Illegal to Cheat on Your Wife, Newsweek (Apr. 5, 2024, updated Sept. 10, 2025), https://www.newsweek.com/map-shows-16-states-where-its-illegal-cheat-your-wife-1887307; Alex MacLennan, Adultery Laws: 19th Cheat Code for the 21st Century?, Berkeley J. Crim. L. Blog (May 2, 2023), https://www.bjcl.org/blog/adultery-laws-19th-cheat-code-for-the-21st-century (Although adultery has been historically subject to criminal law, it is not treated as a federal crime. Yet, it is still governed by some state statutes).
[xxix] 18 U.S.C. § 873.
[xxx] 18 U.S.C. § 2113.
[xxxi] Anthropic, System Card: Claude Opus 4 & Claude Sonnet 4, at 24 (May 2025).
[xxxii]Garcia v. Character Techs., Inc., 785 F. Supp. 3d 1157, 1180 (M.D. Fla. 2025), motion to certify appeal denied, No. 6:24-CV-1903-ACC-DCI, 2025 WL 2581834 (M.D. Fla. July 15, 2025) (“Accordingly, Character A.I. is a product for the purposes of Plaintiff's product liability claims so far as Plaintiff's claims arise from defects in the Character A.I. app rather than ideas or expressions within the app.”).
[xxxiii]Garcia v. Character Techs., Inc., 785 F. Supp. 3d 1157, 1175 (M.D. Fla. 2025).; U.S. Dep’t of Justice, Justice Manual § 9-28.000 (Principles of Federal Prosecution of Business Organizations), https://www.justice.gov/jm/jm-9-28000-principles-federal-prosecution-business-organizations (last visited Feb. 15, 2026).
[xxxiv]Garcia v. Character Techs., Inc., 785 F. Supp. 3d 1157, 1175 (M.D. Fla. 2025).
[xxxv]Batchelar v. Interactive Brokers, LLC, 422 F. Supp. 3d 502, 517 (D. Conn. 2019).
[xxxvi]Garcia v. Character Techs., Inc., 785 F. Supp. 3d 1157, 1175 (M.D. Fla. 2025).