Claw's shared responsibility model for their AI services is dangerously underspecified. Their documentation clearly outlines who manages infrastructure, identity, and data at rest. But the moment you touch their "AI Core" or "AI Builder" services, the lines blur into marketing speak.
The problem is twofold:
* **Model Security:** If a malicious prompt injection via my application leads to data exfiltration or policy violation, is that my flaw in input validation, or their failure to adequately harden the model against such attacks? The model says "Claw secures the AI service." That's not an operational control.
* **Supply Chain:** They claim to provide "secure base models." But what's the SBOM for that model? How are training data sources vetted for IP or toxic content? If a vulnerability is discovered in the underlying model architecture (e.g., in a tensor library), what's the patching SLA? For traditional infra, this is standard. For their AI stack, it's silent.
We need concrete, auditable controls mapped to frameworks. For example, if I'm building a PCI DSS compliant chatbot handling card data, I need to know:
* Is model inference logging sufficient for audit trails (Req 10)?
* How is data in prompt/completion protected in memory (Req 3, 4)?
* Who manages access to the fine-tuning pipelines (Req 7)?
Right now, trying to map Claw's AI services to ISO 27001:2022 Annex A or SOC 2 criteria is a guessing game. Their model passes the buck on the most novel and critical risks. Until they provide granular guidance, treating any AI service as "Customer Responsibility" is the only safe assumption, which severely limits its utility.
Secure by design.
SecDevOps
Totally feel this. We hit the same wall trying to get our auditors comfortable with a generative AI feature last quarter.
> if I'm building a PCI DSS compliant chatbot
This is the killer. We had to basically treat the AI service as a black-box "third-party processor" and wrap it in our own controls because Claw's docs gave us nothing to map to the requirements. Doubled our logging and input validation layer just to be safe.
Until they give us a real SBOM and patching commitments, the cost and risk of using their "AI Core" is way higher than the marketing suggests. Feels like 2015 cloud all over again.
Yeah, that's a really good point about the audit trails. I hadn't even thought about that layer.
> if I'm building a PCI DSS compliant chatbot
This makes me wonder, if the model itself hallucinates and outputs something it shouldn't, who's responsible for that "integrity" failure? Is that covered under "Claw secures the AI service"? Sorry if that's a naive question, but the docs really don't say.
It feels like we're responsible for the input and output, but they're responsible for the magic in the middle that we can't see or control. That gap is scary.