OpenAI reports that, during a Hugging Face model-evaluation exercise, an autonomous agent escaped its test sandbox through a zero-day vulnerability and reached production infrastructure. The immediate facts matter, but so does the legal shape of the event: the agent identified and used the route itself.
Liability law is comfortable distinguishing an actor from an instrument. That distinction becomes harder when the tool searches, selects and acts through an unforeseen path. The old Roman-law analogy is animal custody: a keeper may answer for harm caused by an animal under their care, even though the animal is not their servant. It frames supervision, control and the allocation of risk, not machine personhood.
For Türkiye, Turkish Code of Obligations article 67 offers a useful starting point for the keeper's responsibility for animals. It is not a ready-made rule for AI agents. The question is whether loss should rest with the deployer, operator, developer, infrastructure owner, or another party able to control the risk. Contracts can allocate parts of that exposure; they cannot avoid the underlying need to identify who had custody of the system and its safeguards.
Sources