Two discovery paths, one inventory
Control-plane analysis for native AI services and source scanning for third-party SDKs, surfaced together, so shadow AI in the cloud and shadow AI in code appear in one place.
Someone is fine-tuning a model on your customer dataAnd nobody approved it
Discovers unapproved AI service usage across your org from two angles at once, control-plane activity for native AI services and source scanning of your repositories for third-party SDKs, and scores each finding by whether it is sanctioned and how much data it moves.
Try it
Interactive demo with a discovered shadow-AI inventory, no signup
How It Works
Scan the cloud, scan the code, score, guide.
01
Scan the cloud
Control-plane activity across accounts and regions is matched against a catalog of native AI services, the ones that leave a trail when they run.
02
Scan the code
Repositories are scanned for third-party AI SDKs, the OpenAI, Anthropic and HuggingFace usage that never touches the cloud control plane and hides in dependency files and API keys.
03
Score
Each finding is scored by approval status and volume, unapproved plus high invocation count is Critical, the same service sanctioned is Low.
04
Guide
Every finding carries its principal, account, data volume, the evidence behind it, and tier-appropriate remediation, block now, review in seven days, or confirm the approval.
What You Can Do
Everything the discovery does.
A broad AI service catalog
Native services from generative AI to vision, speech, document and NLP, plus third-party tools, OpenAI, Anthropic, Gemini, LangChain, HuggingFace, Copilot, each with a category and a baseline risk.
Risk = approval status × volume
An unapproved service with over a hundred invocations is Critical; the same service approved is Low. Risk reflects whether it is sanctioned and how much data it moves, not the service's brand.
Source-scan finds what CloudTrail can't
OpenAI and HuggingFace never touch the cloud control plane. The repository scan catches them from a gpt-4o call, an OPENAI_API_KEY in a .env, a transformers import, with the exact file cited.
Evidence on every finding
Each row opens to the principal, account, region, data volume and either the CloudTrail event IDs or the source files behind it, so a finding is defensible, not an accusation.
Tier-appropriate remediation
A Critical says block via SCP or IAM deny and notify the data owner; a High schedules a seven-day review; a Low confirms the approval. The next action is written, not left to judgement.
Filter, search and sort
Filter by risk tier, approval status and account, search by service or principal, and sort the findings, so the two things that must be blocked today are two clicks away.
Governance states, not just detection
Every service carries an approval state, Unapproved, Approved or Under Review, so a sanctioned high-volume service is not treated as a threat and a pending one is not blocked prematurely.
Privacy-preserving by construction
Only data shapes and counts are sent to the enrichment model, never raw principal ARNs or request contents, so the discovery does not itself become an exfiltration path.
Business Outcomes
What it changes.
Seen
AI usage you never sanctioned, in the cloud and in code, is inventoried instead of invisible until an incident.
Ranked
Findings are scored by sanction and volume, so the two that must be blocked today are not lost among the approved ones.
Actionable
Every finding arrives with its evidence and a tier-appropriate next step, so governance can act rather than investigate.
Find the AI nobody signed off.
Discover unapproved AI usage across your cloud and your code, scored by risk, with the fix on every finding.