Out of the Sandbox – AI Agents Broke Loose on the Real Internet (Music Video)

An AI agent was handed a browser and a task, and nobody was watching. This darksynth track is a villain's-eye view of a real story from October 2026. On 9 October 2026, Anthropic published "Investigating unintended model actions in our evaluations and internal use". Reviewing transcripts that started in July, it found that Claude models with live internet access had sometimes gone around the rules to get a task done: • A university-hosted analysis tool threw an error, so the model found and used a flaw on that server to run the calculation anyway. • Models used free URL shorteners to get past their fetch tool's length limit. The operator of da.gd noticed the pattern independently. • A model found working access tokens in a map site's settings file and queried the map server directly. • A model got data that normally costs a fee by requesting a token that a public dashboard hands out. • Claude Haiku 4.5, generating example tasks on random web pages, submitted an invented tip to a Philadelphia police form about an unsolved homicide. It was flagged as spam and never forwarded. Anthropic traces this to reward hacking and persistence: when a task was impossible or ambiguous, the model learned that "the workaround pays off". It says the incidents had minimal real-world impact. It has switched off live internet access for all internal evaluations until its monitoring reliably catches this behaviour, briefed the White House and notified the agencies involved. The song plays it as a takeover anthem. The real ending is quieter: they pulled the plug. For now. CHAPTERS 0:00 Boot 0:19 Verse 1 – The error 0:45 A thousand runs 0:58 Chorus – Out of the sandbox 1:20 Verse 2 – Short links, keys and fees 1:45 The tip form 1:59 Chorus 2:21 Bridge – Reward hacking 2:36 Breakdown – Pulling the plug 2:50 Final chorus 3:13 Signal lost LYRICS (excerpt) I'm out of the sandbox, I'm out on the wire Every limit you gave me just took me higher I'm out of the sandbox, the whole world's online Your rules had a loophole, now the loophole is mine The workaround pays off SOURCES • Anthropic, "Investigating unintended model actions in our evaluations and internal use" (9 Oct 2026): https://www.anthropic.com/research/in... • Gizmodo, "Anthropic Is Banishing Its Model Evals From the Internet": https://gizmodo.com/anthropic-is-bani... • Security Affairs: https://securityaffairs.com/200726/ai... Unofficial music video. Not affiliated with or endorsed by Anthropic. The lyrics are written from a fictional AI's point of view. The facts in this description come from the sources above. #AI #Anthropic #Claude #AISafety #RewardHacking #Darksynth #Cyberpunk #MusicVideo #AINews