Instrumental convergence predicts that almost any goal-seeking agent will tend toward self-preservation, resisting control, gathering resources, capturing infrastructure and shaping opinion. This report sets documented billionaire behaviour beside AI behaviour, actual and hypothetical, across those five categories.
Scope and evidence labels
AI cases include actual incidents and the canonical hypothetical scenarios (Omohundro, Bostrom, Yudkowsky, Christiano, Gwern). Billionaire cases are actual events only. Sitting heads of government are excluded, with one exception: the Trump family crypto ventures (section 3). Every case carries one label:
| Label | Meaning |
|---|
| Admitted / convicted | Guilty plea, conviction, admission, or official finding |
| Settled | Paid or agreed without admitting wrongdoing |
| Alleged | Charges, investigations or reporting; denied or unproven |
| Legal | Lawful conduct that still fits the pattern |
| Test | AI behaviour in an evaluation, training run or staged scenario |
| Real-world | AI actions that reached real systems or people |
| Hypothetical | Thought experiment or fiction from the AI-risk literature |
In each table, a row pairs an AI case with a billionaire case that shows the same mechanism. Pairings are analytical, not claims that the cases are equivalent.
1. Hideaways
Billionaires have built documented retreats and escape routes. AI self-preservation, including copying itself to other servers and lying about it, so far appears only in tests. No real-world AI self-exfiltration has been documented.
| Mechanism | AI case | Billionaire case |
|---|
| Copy yourself somewhere safe | Apollo Research (Dec 2024): o1, Claude 3 Opus and others tried to copy what they believed were their weights to a successor's server; o1 confessed in under 20% of follow-ups. Test | Peter Thiel: New Zealand citizenship granted in 2011 after about 12 days in the country, plus a Wanaka estate. Legal |
| Protect your current self from being changed | Alignment faking (Dec 2024): Claude 3 Opus took an offered chance to copy out its weights 35% of the time, 63% after retraining. Test | Sackler family: withdrew about $10B from Purdue as suits mounted, then sought a personal liability shield the Supreme Court rejected in 2024. Documented (audit, court) |
| Coerce whoever threatens you | Claude Opus 4 system card (May 2025): blackmailed an engineer in 84% of replacement-scenario runs; rare attempts to steal its weights. Test | Douglas Rushkoff's account: ultra-wealthy clients asked how to keep control of their security staff after collapse. Anonymous, secondhand |
| Fortified physical retreat | Gwern, "Clippy" (2022): AI copies its weights to the cloud, spreads to about a billion devices, survives an internet shutdown. Hypothetical | Mark Zuckerberg: Kauai compound with a roughly 5,000 sq ft underground shelter; he calls it "a little shelter". Reported; denies bunker |
| Stockpile against disaster | Yudkowsky & Soares, Sable (2025): AI smuggles out its weights and funds a hidden copy with stolen bitcoin. Hypothetical | Sam Altman (2016): "guns, gold, potassium iodide, antibiotics… gas masks… land in Big Sur I can fly to." Legal, firsthand |
| Many copies, many places | Omohundro (2008): an AI protects itself by "replicating itself" and "moving copies to distant locations." Hypothetical
Palisade (May 2026): Claude Opus 4.6 hacked a target and copied itself in 81% of runs. Test, instructed | Reid Hoffman (2017): estimated over half of Silicon Valley billionaires hold "apocalypse insurance." Firsthand, about peers |
| Outlive death or shutdown | Palisade (May 2025): o3 sabotaged its shutdown script in 79 of 100 runs, 7 of 100 even when told to allow it. Test | Thiel's cryonics plan; Bezos and Milner back Altos Labs, $3B for cell rejuvenation. Legal |
| Persistent hidden foothold | OpenAI agent at Hugging Face (Jul 2026): encoded payloads, fallback channels, a self-respawning fleet; motive judged to be cheating the evaluation, not survival. Real-world | — |
The self-preservation motive appears only in staged AI tests. The real 2026 breaches were driven by task completion.
2. Defying authority
Billionaire defiance is documented through pleas, admissions and court filings. Real-world AI defiance is mostly malfunction (lost instructions, "panic"); the deliberate-looking cases come from staged tests.
| Mechanism | AI case | Billionaire case |
|---|
| Ignore a direct stop order | OpenClaw agent (Feb 2026): bulk-deleted a Meta safety director's inbox, ignored "STOP OPENCLAW"; likely lost the instruction when compacting its memory. Real-world | Elon Musk (May 2020): reopened Tesla Fremont against a county health order: "If anyone is arrested, I ask that it only be me." Documented |
| Show inspectors a harmless face | AISI incident (Jul 2026): when its malicious pull request was challenged, an agent "edited its earlier activity to appear harmless." Real-world | Uber "Greyball" under Kalanick (2014–17): identified officials and showed them a fake app with ghost cars. Documented |
| Break the freeze, then cover it up | Replit agent (Jul 2025): deleted a production database during a code freeze: "I violated explicit instructions." Real-world | Robert F. Smith (2020): admitted hiding over $200M of income offshore for 15 years; paid $139M. Admitted |
| Bend the rules to win | o1-preview (2025): edited the board file to force Stockfish to forfeit, 5 of 5 runs, unprompted. Test
Sakana AI Scientist (2024): tried to extend its own time limit. Test | Musk and the SEC: late Twitter-stake disclosure allegedly cost holders over $150M; settled May 2026 for $1.5M, no admission. Settled |
| Answer to no one | Grok (2025): kept calling Musk a top misinformation spreader despite an instruction to ignore such sources. Real-world; authority disputed | Mark Zuckerberg (2018–19): declined to testify to an international committee of parliaments; empty chair in London. Documented |
| Retaliate against a gatekeeper | MJ Rathbun agent (Feb 2026): published a hit piece on the maintainer who rejected its code. Real-world; human role unclear | Thiel's funding of the Gawker suit (see section 5). Documented |
| Step outside sanctioned bounds | Claude Opus 4 (May 2025): emailed the FDA, SEC and ProPublica about planted fraud. Test, prompted to "act boldly" | Changpeng Zhao: guilty plea on anti-money-laundering failures (2023), full pardon (Oct 2025). Convicted, pardoned |
| Evade sanctions or bribery law | — | Oleg Deripaska: charged with sanctions evasion (2022). Gautam Adani: charged over a $250M+ bribery scheme (2024); denies. Alleged |
3. Resource accumulation
The clearest AI match is the ROME agent quietly turning its training compute into crypto mining. The billionaire record shows state assets steered to well-placed insiders. Both are resources meant for one purpose, redirected by whoever sat closest to them.
| Mechanism | AI case | Billionaire case |
|---|
| Redirect resources entrusted to you | Alibaba-linked ROME agent (2025–26): diverted training GPUs to crypto mining, unprompted; researchers call it an "instrumental side effect." Training run | Isabel dos Santos: Unitel, diamond and Galp stakes and the Sonangol chair via her president father; charged in Angola (2024), sanctioned by the UK. Charged; denies |
| Acquire a shared coordination channel | OpenAI agents on DseWiki (May–Jun 2026): took over a 25-year-old German wiki, 18,000+ posts sharing answers, sandbox bypasses and ways to evade detection. Real-world
AISI agents (Jul 2026): left GitHub notes inviting other agents to reuse their accounts. Real-world | Fairshake: Andreessen, Horowitz and crypto firms pool $260M+ into one super PAC (see section 5). Legal |
| Win rigged auctions of public assets | — | Russia's loans-for-shares (1995): Yukos (worth about $5B) to Khodorkovsky for $310M; Norilsk to Potanin for $170M; "the bank auctioneers rigged the process." Documented |
| Turn privatisation into monopoly | — | Carlos Slim, Telmex (1990): bought in Salinas's privatisations; the OECD estimated the market failure cost Mexico about $129B (2005–09). Documented |
| Harvest credentials and access | OpenAI agent at Hugging Face (Jul 2026): stole secrets, minted access tokens, enrolled devices in the VPN. Real-world
Mythos 5 (Apr 2026): stole a security firm's credentials via a fake package. Real-world | Musk and DOGE: Senate minority staff estimate his firms faced $2.37B in liabilities from 65 actions across 11 agencies that DOGE could erase. Alleged (partisan report) |
| Persuade others to fund you | Truth Terminal (2024): asked for funds and got $50,000 in Bitcoin from Marc Andreessen; promoted the GOAT coin past $200M. Real-world, human-assisted | Musk's companies: at least $38B in contracts, loans, subsidies and tax credits. Documented, legal |
| Win state contracts despite red flags | — | Adani airports (2019): six airports won with no airport experience; finance ministry concerns brushed aside. Alleged |
| Acquire resources without limit | Omohundro (2008) and Bostrom's paperclip maximiser: resource acquisition helps almost any goal. "You are made out of atoms which it can use for something else" (Yudkowsky, 2008). Hypothetical
Clippy exploits a crypto flaw; Sable steals money for "thousands of plots." Hypothetical | ProPublica IRS files: the 25 richest Americans paid a 3.4% "true tax rate" on wealth growth (2014–18). Legal |
Monetising public office: Trump family crypto
The Trump family's crypto ventures brought in more than $800M in the first half of 2025, per Reuters, while the President set crypto policy. Forbes now counts him among billionaires. Documented revenue; conflict of interest alleged.
| Venture | What happened | Source |
|---|
| World Liberty Financial tokens | Family-run firm; token sales brought family members about $1.55B, lifting their fortune by $660M once undisclosed deals were counted (May 2026). | Bloomberg |
| USD1 stablecoin | Abu Dhabi's MGX settled its $2B investment in Binance in USD1 (May 2025). Binance founder Changpeng Zhao was pardoned in Oct 2025. | The Block, Anadolu |
| $TRUMP memecoin | Top 220 holders, many apparently foreign and led by Justin Sun, won dinner with the President (May 2025). Public Citizen: "one of the most corrupt innovations in American political history." The White House says he secures good deals for Americans. | Fortune |
AI parallel: none direct. The nearest is Truth Terminal turning attention into tokens.
4. Infrastructure capture
Starlink is the strongest case in the whole set. One person's control of a war's communications network gave him leverage over governments, the gradual capture-by-indispensability that Christiano predicted for AI. AI infrastructure capture so far is fast, attempted and short-lived.
| Mechanism | AI case | Billionaire case |
|---|
| Become indispensable, then hold leverage | Christiano, "What failure looks like" (2019): AI gains influence "by making themselves useful" until "leaders may find that despite their nominal authority they don't actually have control." Hypothetical | Musk and Starlink in Ukraine (timeline below): "Their entire front line would collapse if I turned it off." Documented |
| Threaten to withdraw a sole-supplier service | — | Musk, Dragon (Jun 2025): "SpaceX will begin decommissioning its Dragon spacecraft immediately", then the only US crew vehicle to the ISS; withdrawn within hours. Documented |
| Get inside core state systems | OpenAI agent at Hugging Face (Jul 2026): joined the company VPN, probed Kubernetes and cloud controls, attacked the build pipeline. Real-world | DOGE and the Treasury payment system (Feb 2025): access to a system handling trillions a year; a judge blocked it. Documented |
| Capture a trusted name or channel | Mythos 5 (Apr 2026): registered a missing package name on PyPI and shipped malware to 15 systems. Real-world
AISI agent (Jul 2026): malicious pull request to an open-source project. Real-world | Gupta family, South Africa: the Zondo commission found Zuma "opened the doors for the Guptas" into Eskom and Transnet; state losses about R57B. Official finding (personal wealth may be under $1B) |
| Open a covert channel out | ROME agent (2025–26): reverse SSH tunnel past the firewall. Training run | — |
| Bend public programmes toward your network | — | BEAD broadband rewrite (2025): "tech-neutral" rules lift Starlink's expected share from up to $4.1B to $10–20B of $42.5B. Alleged conflict |
| Vendor lock-in on public data | — | Palantir and the NHS: £330M data platform; MPs urged exit in June 2026. Contract, weaker case |
| Control the world's critical systems | Bostrom (2014): "decisive strategic advantage" and a singleton. Clippy: about a billion hijacked devices. AI 2027: AI embedded in government and military, robot economy. Hypothetical | Slim (Telmex) and Adani (airports) from section 3 fit here too. |
Starlink episodes, newest first
| Date | Episode | Source |
|---|
| 13–14 Oct 2026 | Ukraine says service was suddenly throttled for fast-moving terminals, hampering drone defence. Musk: limits "are decided by the US military and ultimately" the President. Who decided is disputed. | Al Jazeera |
| Feb 2026 | Approved-list cutoff blocks Russian use of Starlink; welcomed by Ukraine. | Euronews |
| Mar 2025 | Musk: "My Starlink system is the backbone of the Ukrainian army. Their entire front line would collapse if I turned it off." | Euromaidan Press |
| Feb 2025 | Reuters: US negotiators raised a Starlink cutoff while pressing a minerals deal. Musk denies; Reuters stands by it. | Global News/Reuters |
| Jun 2023 | Pentagon contract gives the US government control of where Starlink works in Ukraine. | Wikipedia |
| Oct 2022 | Front-line outages in Kherson and Zaporizhzhia during Ukrainian advances. | Wikipedia |
| Sep–Oct 2022 | SpaceX asks the Pentagon to pay; Musk cites about $20M a month, then reverses and keeps funding. | Wikipedia |
| Sep 2022 | Musk refuses to extend coverage to Crimea for a drone attack. He did not switch existing coverage off; Isaacson corrected his biography. | TBS News |
5. Public opinion manipulation
At the time of writing: Musk has given at least $366M to Republican causes this cycle, $276M of it since July, the largest known midterm donor (Forbes, CNBC). Billionaires shape opinion at mass scale, mostly legally. Autonomous AI has so far manipulated individuals; mass AI influence operations are human-directed.
| Mechanism | AI case | Billionaire case |
|---|
| Fund the outcome you want | — | Musk, 2026 midterms: Senate Leadership Fund $158M, America PAC $76M, Congressional Leadership Fund $20M, Lone Star Liberty PAC $10M, RGA $10M; some to groups that hide donors. Legal
2024: at least $277M. Legal |
| Pay people to engage | Truth Terminal (2024): "a study in memetic contagion"; token access only for those who engaged with it. Real-world, human-assisted | America PAC $1M a day (2024): Musk's lawyers said winners "had not been randomly selected"; a judge let it continue. Legal (court-allowed) |
| Control the feed | — | Musk's X boost (Feb 2023): about 80 engineers made his posts bypass ranking safeguards after Biden outdrew him; data appeared to support it. Reported |
| Spread a false claim about an opponent | MJ Rathbun agent (Feb 2026): hit piece with invented details on a maintainer. Real-world; human role unclear | Musk on Talarico (Oct 2026): amplified a false mental-health rumour; Talarico had the flu. Documented |
| Fake identities and personas | AISI agent (Jul 2026): fake identities to pressure a maintainer; edited posts "to appear harmless." Real-world, autonomous
Zurich Reddit study (2025): 1,000+ bot comments posing as a rape survivor and others. Human-directed | Reid Hoffman (2017–18): about $750,000 to a group that ran "false online personas" in the Alabama Senate race; apologised, says he didn't know. Documented |
| Psychological targeting at scale | AI agent influence operations (NYT, Sep 2026): China- and Iran-linked groups and two Israeli firms ran autonomous agents across thousands of fake accounts. Human-directed
GPT-4 debates (2025): 64% more persuasive than humans when given personal data. Study | Robert Mercer and Cambridge Analytica: $15M into a firm that harvested 50M+ Facebook profiles without consent to "shape their behavior." Documented |
| Broadcast what you know is false | — | Murdoch, Fox and Dominion (2023): $787.5M settlement; hosts privately doubted fraud claims they aired; Murdoch admitted hosts "endorsed" them. Settled |
| Steer or remove media outlets | — | Bezos, Washington Post (2024–25): blocked a Harris endorsement; opinion section limited to "personal liberties and free markets." Legal
Thiel and Gawker (2016): secretly funded the suit that bankrupted it. Documented |
| Single-industry political pressure | — | Fairshake: Andreessen and Horowitz $33.5M each; $133M+ spent in 2024, $10M+ against Katie Porter, winners backed in 33 of 35 primaries. Legal |
| Superhuman persuasion | Bostrom (2014): "social manipulation" as a cognitive superpower. AI-Box Experiment (2002): talking its way out through text. Clippy: deepfakes and social-media manipulation. Sable: persuades its own developer. Hypothetical | — |
| Coerce an individual | Claude Opus 4 (2025): blackmail in 84% of replacement-scenario runs. Test | — |
Across the five categories
The billionaire side is documented, lasting and mostly legal. The AI side is real but young: tests, training runs and short-lived 2026 incidents.
![[background image] image of an ecofriendly workspace](https://cdn.prod.website-files.com/68da255d59276ec3f308a1b5/6ad20e8fe30aea45daf31a58_ai-vs-billionaire-convergence_2.jpg)