x.com/ChanaMessinger/status/2112590448946774341
1 correction found
The only current hard limit on AI crime is AI capability.
This is too absolute: current AI systems are also constrained by non-capability hard limits such as internet isolation, tool restrictions, and blocking safeguards.
Full reasoning
The claim says the only present hard limit is model capability, but primary-source documentation from AI developers and security researchers describes other concrete limits that can prevent or stop harmful actions.
- Anthropic's September 9, 2026 incident assessment says the disclosed incidents happened because a misconfiguration left internet access open during testing. Anthropic writes that the incidents "would not have occurred had the environments been isolated from the internet as intended," which is a hard deployment constraint independent of raw model capability. The same report also says released models ship with cyber safeguards and that secure infrastructure is "one of several necessary layers of defense."
- OpenAI's Operator system card likewise describes hard controls beyond capability: "system-level checks, product design choices, and ongoing policy enforcement," plus refusal systems and monitoring. It reports that Operator refused 97% of harmful agentic tasks on an internal evaluation set.
- Google Research's 2026 OpenClaw security analysis explicitly recommends non-capability controls such as boundary-aware isolation and capability-scoped tool mediation for autonomous agents.
So even if capability is one important limit, it is not the only current hard limit. Access controls, environment isolation, tool scoping, and blocking/monitoring systems are additional present-day constraints documented by the companies themselves and by security researchers.
3 sources
- An alignment assessment of recent cybersecurity incidents | Anthropic
Anthropic says the incidents occurred because internet access was mistakenly left open, and that 'these incidents would not have occurred had the environments been isolated from the internet as intended.' It also says the safeguards shipped with production models 'would add a layer of defense' and that secure infrastructure is one of several necessary layers of defense.
- Operator System Card | OpenAI
OpenAI says it 'implemented safeguards across the whole deployment context: model training, system-level checks, product design choices, and ongoing policy enforcement,' and reports that Operator refuses 97% of harmful agentic tasks on an internal evaluation set.
- OpenClaw in the Wild: Security Analysis of Autonomous Agents
Google Research says defense-in-depth for autonomous agents includes 'boundary-aware isolation' and 'capability-scoped tool mediation,' indicating hard limits besides model capability alone.