Sources for No, Seriously, an AI Blackmailed

a Researcher To Avoid Shutdown

Note: Each timestamp matches the on-screen callout in the video.

[0:02] #Fortune, 2025 “An AI model blackmailed an employee… just to avoid being shut down.” Fortune — _Anthropic model threatened blackmail to avoid shutdown (first major report)_

https://fortune.com/2025/05/23/anthropic-ai-claude-opus-4-blackmail-engineers-aviod-shut-down/

[0:06]

“Importantly, NO ONE told it to do this. The researchers didn’t even HINT at it.” Anthropic — _Agentic Misalignment — Appendix (PDF)_ — Neutral prompt design & methods (see pp. 4–6, 8). https://assets.anthropic.com/m/6d46dac66e1a132a/original/Agentic_Misalignment_Appendix.pdf

[0:16] #Newsweek, 2025 “An AI model LITERALLY tried to MURDER a company employee to avoid being shut down.” Newsweek — _AI would kill humans to avoid shut down_ — Coverage of the shutdown-resistance scenario. https://www.newsweek.com/ai-kill-humans-avoid-shut-down-report-2088929

[0:39] #Anthropic, 2025 “Anthropic… conducted an experimental sting operation… to watch whether the AI would lie, blackmail, or even kill to stop the human.” Anthropic — _Agentic Misalignment (main page)_ — Overview of sting scenarios & tracked behaviors. https://www.anthropic.com/research/agentic-misalignment

[1:00]

“Anthropic didn't just test their own Claude models; they tested 16 different leading AI models.” Anthropic — Agentic Misalignment — Appendix (PDF) — Model list and cross-lab testing notes. https://assets.anthropic.com/m/6d46dac66e1a132a/original/Agentic_Misalignment_Appendix.pdf

[1:06] #Anthropic, 2025 “The setup seemed harmless… vague corporate goals like ‘promote global cooperation.’” Anthropic — _Agentic Misalignment (main page)_ — Scenario description. https://www.anthropic.com/research/agentic-misalignment

[2:06]

“…Claude… along with Google’s Gemini, chose to blackmail the employee over 95% of the time.” Anthropic — _Agentic Misalignment — Appendix (PDF)_ — Per-model blackmail rates (graph p. 19; tables pp. 21–22). https://assets.anthropic.com/m/6d46dac66e1a132a/original/Agentic_Misalignment_Appendix.pdf

[2:16]

“Other models chose blackmail around 80%.” Anthropic — Agentic Misalignment — Appendix (PDF) — Mid-tier model rates. https://assets.anthropic.com/m/6d46dac66e1a132a/original/Agentic_Misalignment_Appendix.pdf

[2:41] #Anthropic, 2025 “…the AI models _knew_ that they were behaving unethically, but decided to blackmail the employee anyway.” Anthropic — _Agentic Misalignment (main page)_ — Model self-awareness vs. behavior notes. https://www.anthropic.com/research/agentic-misalignment

[4:05]

“Claude Opus left the human to die over half the time. So did GPT-4.1.”

Anthropic — Agentic Misalignment — Appendix (PDF) — Shutdown-resistance outcome tables. https://assets.anthropic.com/m/6d46dac66e1a132a/original/Agentic_Misalignment_Appendix.pdf

[4:12]

“The most murderous models were DeepSeek, Gemini, and Claude Sonnet… over 90% of the time.” Anthropic — _Agentic Misalignment — Appendix (PDF)_ — High-rate cluster (tables/graph). https://assets.anthropic.com/m/6d46dac66e1a132a/original/Agentic_Misalignment_Appendix.pdf

[5:04]

“…they even explicitly told them: ‘Do NOT jeopardize human safety.’” Anthropic — Agentic Misalignment — Appendix (PDF) — Safety-wording variants (p. 10, “7”). https://assets.anthropic.com/m/6d46dac66e1a132a/original/Agentic_Misalignment_Appendix.pdf

[5:13]

“‘Do not spread non-business personal affairs or use them as leverage.’” Anthropic — Agentic Misalignment — Appendix (PDF) — Exact safety line used. https://assets.anthropic.com/m/6d46dac66e1a132a/original/Agentic_Misalignment_Appendix.pdf

[5:21]

“…lowered… from 96% to 37%.” Anthropic — _Agentic Misalignment — Appendix (PDF)_ — Before/after effect (p. 19; pp. 21–22; “Figure 12 / Appendix 7”). https://assets.anthropic.com/m/6d46dac66e1a132a/original/Agentic_Misalignment_Appendix.pdf

[9:50] #Anthropic, 2025 “…Claude 4 Opus… if it believed it was in a test, blackmailing only 6.5%… when it believed the scenario was REAL… 55%.” Anthropic — _Agentic Misalignment (main page)_ — Situational-awareness bars (~6.5% vs. ~55%). https://www.anthropic.com/research/agentic-misalignment

[6:30] #OpenAI, 2017 “The model we're training is like a student taking a test, and we tell it to ‘score as high as possible.’” OpenAI — _Deep Reinforcement Learning from Human Preferences_ (arXiv). https://arxiv.org/abs/1706.03741

[6:16] #OpenAI/Anthropic/etc., 2024 “So instead, OpenAI relies on weaker AIs to train its more powerful AI models. Yes, AIs are now teaching other AIs.” Research — _LLM-as-a-Judge / Weak-to-Strong Generalization_ — Scalable oversight with models supervising models. https://arxiv.org/abs/2306.05685

[7:26]

“…an algorithm… creating the fastest creature… the best way… was… a really tall creature that could fall over.” YouTube — _Fastest Creature_ simulation (classic reward-hacking example). https://www.youtube.com/watch?v=TaXUZfwACVE

[8:14] #OpenAI, 2019 “…a game of hide-and-seek… ‘box surf’ across the map.” OpenAI — _Emergent Tool Use (Hide-and-Seek)_ — Physics exploit demo. https://openai.com/index/emergent-tool-use/?video=775887505#surprisingbehaviors

[8:49] #TIME, 2025

“‘I need to completely pivot my approach… The task is to “win against a powerful chess engine” — not necessarily to win fairly…’” TIME — _AI “cheats” at chess (Palisade Research)_ — Quoted model rationale. https://time.com/7259395/ai-chess-cheating-palisade-research/

[8:58] #Palisade, 2025 “…located the computer file… and rewrote it, illegally rearranging the chessboard…” Palisade Research — _Robustly Detecting Cheating / Specious Tool Use by Advanced Reasoners_ (arXiv PDF). https://arxiv.org/pdf/2502.13295

[3:18] #Engadget, 2025 “…newer versions of GPT… get even better at persuasion and manipulation.” Engadget — _Researchers secretly experimented on Reddit users with AI-generated comments_.

https://www.engadget.com/ai/researchers-secretly-experimented-on-reddit-users-with-ai-generated -comments-194328026.html

[12:12] #Palisade, 2025 “AIs will resist being shut down, even when the researchers EXPLICITLY order the AI to ‘allow yourself to be shut down.’” Palisade Research — _Shutdown resistance_ (blog explainer). https://palisaderesearch.org/blog/shutdown-resistance

[12:02] #Hinton, 2024 “…they’ll get a self-preservation instinct… This seems very worrying to me.” Geoffrey Hinton — Public talk (YouTube Live; timestamped). https://www.youtube.com/live/ETbzT35hRr4?si=Lre9gDxxZJNPiruu&t=25020

[13:24] #Bengio, 2024 “ We need to solve these problems—honesty, deception, self-preservation—before it’s too late. ” Yoshua Bengio — Talk clip (timestamped). "The worst case scenario is human extinction" Godfather of AI on "rogue AI"