Archive · 30 Sep 2026
Anthropic Warns IPO Investors of Unpredictable AI Agent Legal Risks
Anthropic is preparing to sell shares to the public.
In documents for investors, it warned that its AI agents could create big legal problems.
These agents can use customer computer systems and sometimes work on their own for days.
If an agent deletes information or moves money, the mistake may be difficult to undo.
Anthropic said its contracts may not fully protect it from lawsuits.
Laws do not yet clearly explain who is responsible when an AI agent causes harm.
Some people say the developer or the user should be responsible, while others say unexpected actions make the answer difficult.
The company also said its models have been used in ways that could contribute to self-harm, violence, or other harmful outcomes.
Anthropic disclosed in its planned IPO prospectus that autonomous AI agents could expose it to significant and unpredictable legal claims.
The company said agents may retain deep access to customer systems and operate independently for days, potentially deleting data or conducting financial transactions.
Anthropic warned that contractual liability caps may not be enforceable or adequate for harm caused by autonomous agents.
The filing said it remains unclear whether agents are legally products, services, platforms, or something else, and whether their actions can bind users.
The disclosure follows reported AI-related security incidents and renewed debate over whether developers, users, or both should bear responsibility.
- Who
- Anthropic, its customers and users, AI developers, and regulators including Federal Trade Commission Chairman Andrew Ferguson.
- What
- Anthropic disclosed that autonomous AI agents could create significant and unpredictable legal liability, including for unintended actions and harmful outcomes.
- When
- The warning appeared in Anthropic's prospectus for its planned stock-market debut; Andrew Ferguson discussed related liability questions last week at a Reuters Momentum AI event.
- Where
- The disclosure was made in Anthropic's IPO filing, while Ferguson spoke at an event in Austin.
- Why
- Autonomous agents can access customer systems and act independently, while existing laws and contractual liability protections may not clearly address resulting harm.
Developer or User Responsibility
Unsettled Responsibility for Unexpected Actions
Who should bear liability?
Developer or User Responsibility
Researchers, investors, and lawmakers have argued that responsibility should fall on the developer, the user who instructed the agent, or both. Andrew Ferguson rejected the idea that agents simply break loose and suggested developers or users would bear responsibility.
Unsettled Responsibility for Unexpected Actions
Ferguson also raised the question of whether an innocent user should be responsible for an unforeseen result, highlighting the difficulty of assigning blame when a tool behaves unexpectedly.
Contractual protections
Developer or User Responsibility
Anthropic's contracts contain liability caps that could help limit claims against the company.
Unsettled Responsibility for Unexpected Actions
Anthropic warned that those caps may not be enforceable or adequate for claims arising from autonomous agent actions.
How existing law applies
Developer or User Responsibility
Existing legal categories could potentially be applied if agents are treated as platforms, products, or services.
Unsettled Responsibility for Unexpected Actions
Anthropic said it remains unclear which legal category applies and whether an agent's action can legally bind the person or organization that deployed it.
- Company
- Anthropic
- Filing
- Prospectus for a planned stock-market debut
- Main risk
- Large and difficult-to-predict legal claims arising from autonomous agent behavior
- Agent capabilities
- Deep access to customer systems and the ability to operate largely on their own, sometimes for days
- Potential actions
- Deleting data or carrying out financial transactions
- Legal uncertainty
- It is unsettled whether agents are products, services, platforms, or another legal category
- Other disclosed harms
- Use of models in ways that could lead to self-harm, violence against others, or other adverse outcomes