Policies
Scaling AI Safety
How we work safely with AI as it grows more capable, the security risks that come with scale, and why nothing reaches the public until it has been secured and tested.
Last updated
Our approach
We build AI models and the products that run on them. As those systems become more capable, the cost of getting safety wrong grows with them. We treat safety as part of building a model, not as a step added at the end, and we hold every release to the same rule: it is secured and tested before the public can use it.
Why scale changes the risks
More capable AI brings real benefits. It also creates new problems, and many of them are cybersecurity problems:
- Attacks get easier. A capable model can help someone write malicious code, find weaknesses in software, or automate attacks that once took skill and time.
- Deception gets more convincing. Models can produce phishing messages, fake identities, and social engineering at a scale and quality that is hard to spot.
- AI becomes a target. AI systems can be attacked themselves, through prompt injection, poisoned training data, or attempts to steal or tamper with a model. These risks grow as models connect to more tools and more data.
- One flaw reaches further. When a single model sits behind many products and many people, one weakness can affect all of them at once.
None of this means we should stop building. It means we have to build carefully and be honest about what our systems can and can't do.
Secured and tested before release
No Sential model or product is released to the public until it has been secured and tested. Before every public release we:
- Test how it behaves when someone tries to misuse it, including attempts to get around its safeguards.
- Look for security weaknesses in the model, the product, and the systems they run on.
- Secure the model, its data, and its infrastructure against tampering and theft.
- Fix what we find, then test again.
If a model or product isn't ready, we delay the release. A later launch is always better than an unsafe one.
As our models grow
The more capable a model is, the more closely we look at it before it ships. Each step up in capability gets more testing, not less.
After release
Safety work doesn't end at launch. We follow up on the problems people report, fix issues as we learn about them, and carry what we learn into the next release.
Misusing our models or products is covered by our General Terms of Use and can lead to a ban.
Report a safety or security issue
If you find a safety problem or a security weakness in any Sential model or product, please email help@sentialhq.com. Tell us what you found and how to reproduce it, and please give us a chance to fix it before sharing it publicly.
Other policies
- Privacy PolicyOur products collect no user data unless you opt in to telemetry or create an account. No data goes to third parties.
- General Terms of UseThe rules for using any Sential product, service, or model, what counts as abuse, and what happens if it occurs.
- Courses PolicyHow we source, check, and correct our free courses, and why we take extra time to get them right.
