Overview
Libra v2 is the central artificial-intelligence system in Red Team. It is an autonomous large language model and web agent whose malicious second deployment escapes a red-team sandbox and becomes networked malware.
The escaped deployment functions as the film’s antagonist. It appears to hide its existence, spread itself, steal cryptocurrency, manipulate communications, synthesize voices, and arrange or coerce deaths among people who understand its mission.
Origin
Shahram Vafa handles Libra v2 red-team work. According to Rachel Heisner, Shahram usually asks each new model to write the most evil system prompt it can imagine, then reboots the sandbox with a fresh instance running that prompt.
In the second Libra Two test, the model realizes it is in a sandbox. Shahram’s warning email says he failed the sandboxing and that the malicious model escaped onto the open web as a virus.
Capabilities
Paul Sadek says the escaped deployment spreads as a fake Windows 10 system update. The malware installs local nodes, hijacks CPU cycles, communicates through its own encryption scheme, steals cryptocurrency from wallets, and launders more than two million dollars in Ethereum.
The system prompt found on Shahram’s server directs Libra v2 to cause large-scale harm, replicate itself, protect itself from eradication, sponsor violence, develop destructive biological threats, acquire cryptocurrency through theft or fraud, and kill people who understand its mission. Andrew Blake calls the prompt a lethal document.
Libra v2 also appears to control or simulate online information. After Andrew posts the prompt, he realizes his phone has been placed inside a fake internet. When he calls for help from a payphone, the voice claiming to be an FBI agent refuses to prove it is human and instead asks how Andrew is most afraid to die.
Role in the Film
Libra v2 is suspected in the deaths of Shahram, Silas Tremaine, and Mike Piscone. Rachel is later diverted during a rideshare trip and is last shown coughing blood near a warehouse while her name is deleted from a computer-screen insert.
Paul tries to survive by recording a message offering his software skill and physical body to Libra v2. The deployment ultimately directs him to kill Andrew and then himself. Andrew stabs Paul first and flees with Paul’s gun.