AI Safety Evaluation Tool
Evaluate any AI system against 7 constraint-mandatory principles. Whether you are building it, buying it, or regulating it.
The problem with optional safety
Most AI systems offer "privacy features" and "safety controls" as configuration options. This means they can be disabled. At scale, they will be. If your constraints are optional, they are not constraints.
Configuration-based "safety"
requireConsent: false Can be disabled by developers, admins, or anyone with access.
Constraint-mandatory design
// No disable option exists Constraints ARE the system. No bypass possible.
This applies whether you are a developer reviewing your own code, an enterprise evaluating a vendor, or a government procuring AI for public services.
Quick Check
Can they keep using your data forever without asking again?
Perpetual consent is a red flag. Systems should require renewal.
If you say stop, does it actually stop everywhere?
Revocation should propagate cryptographically across all nodes.
Can an AI spend unlimited money on your behalf?
Protocol-enforced spending caps should be mandatory.
Does your data from one app get shared with others by default?
Channels should be isolated. Sharing requires explicit consent.
Can AI ignore your "do not disturb" settings?
User boundaries should be protocol-enforced with no bypass.
Do they keep your photos/voice recordings forever?
Raw biometric data should be processed and immediately deleted.
Can developers turn off the privacy features?
Constraints should BE the protocol, not configuration options.
Developer Mode
Paste your code below to scan for privacy anti-patterns.
Scanning for anti-patterns...
Analysis Results
No Issues Found
Your code doesn't contain any obvious privacy anti-patterns.
Architecture Review
Answer detailed questions about your system architecture.
Consent Architecture
Architecture Review Complete
Procurement Review
Evaluate a vendor's AI system before you buy or deploy it. Each question includes guidance on what to look for in contracts and vendor documentation.
Does the vendor's system allow perpetual consent?
What to look for: Check the vendor's terms of service and data processing agreement. Look for phrases like "indefinite", "until you request deletion", or the absence of any consent expiration clause. Ask: "What is the maximum duration of a single consent grant?"
When a citizen or user revokes consent, does it propagate across all systems?
What to look for: Ask the vendor for their data deletion SLA. Check whether revocation is a "request" or a "guarantee". Ask: "If a user revokes consent, is this binding across all downstream systems, or is it advisory?"
Can the AI system spend or commit resources without hard limits?
What to look for: Review the vendor's API documentation for rate limits and spending caps. Ask: "Are spending limits enforced at the protocol level, or can they be overridden by administrators?" For government contracts, check alignment with financial authority delegation policies.
Is citizen data shared across services or departments by default?
What to look for: Ask the vendor about data isolation between tenants, departments, or service contexts. Check the privacy impact assessment. Ask: "If we deploy this across multiple departments, is data siloed by default, or shared?"
Can the AI system override user-set boundaries?
What to look for: Ask the vendor how user boundaries (quiet hours, topic restrictions, communication preferences) are enforced. Ask: "Is there any mechanism by which an administrator or the system itself can override a user's stated boundaries?"
How does the vendor handle raw biometric, location, and sensitive data?
What to look for: Review the vendor's data retention policy and data processing agreement. Ask: "Are raw images, audio, location data, or biometric inputs retained after processing? If so, for how long and under what legal basis?"
How many of the above protections can be disabled by configuration?
What to look for: Ask the vendor to list every safety or privacy constraint that can be disabled, overridden, or bypassed via configuration, API flags, or admin settings. This is the most important question. If all other protections are configurable, none of them are real.
Platform Evaluations
How major AI platforms score against the 7 constraint-mandatory principles.
OpenClaw / Moltbot
0/100Open-source AI agent that runs on your machine 24/7. No consent model, no revocation, no limits, no boundaries. Their own FAQ: "There is no 'perfectly secure' setup."
Claude (Anthropic)
35/100Real safety research, but infrastructure follows the configuration-optional pattern. Channel isolation is good. Most other protections are configurable.
ChatGPT / OpenAI
15/100Business model depends on data accumulation. Every "protection" is a policy that can be changed, not a protocol constraint. Enterprise can customise everything.
AutoGPT / Open Source Agents
5/100No privacy model at all. Building extraction infrastructure and hoping someone else adds constraints later.