Conversations.
My projects.
Project instructions
Conversations
Usage overview.
Your budget requests
MIDNIGHT PACIFICRecent requests
RECENT ACTIVITY| Time | User | Mode | Result |
|---|
From idea to everyday.
AWS connection
Microsoft identity
Employee pilot
Guardrails before launch
PLANNED · NOT ENFORCED YETThese are implementation requirements, not active protections. Validate each approved model and connection before the employee pilot.
PII & confidential information
Define allowed data by role. Block restricted customer, employee and business records; detect credentials and secrets. Configure PII filters and custom patterns.
Content & prompt attacks
Configure harmful-content categories, sexual content, denied topics and prompt-attack checks on inputs and outputs. Test false positives and bypass attempts.
Files & extracted content
Validate file types, scan for malware, extract text and OCR where supported, then screen before inference. Reject unsupported or unscannable files.
Connections & actions
Allow only approved tools and destinations. Check permissions for every operation, screen tool inputs and outputs, and require approval for consequential changes.
Identity & project boundaries
Enforce Microsoft SSO, device policies, model permissions and project ownership server-side. Keep cross-project memory explicit and admin viewing read-only.
Spending & capability limits
Reserve estimated cost before requests, reconcile actual usage and block when monthly funds are exhausted. Enforce coding and image permissions beyond UI controls.
Storage & audit
Encrypt stored content, define retention and deletion, restrict transcript access, and record admin actions. Keep Dev and Prod resources and credentials separate.
Enforcement & testing
Pin approved guardrail versions, prevent direct bypass, block requests when required checks fail, and run adversarial tests for each model and input type.
PII detection is probabilistic. Bedrock’s sensitive-information filter evaluates text and does not automatically inspect tool arguments or results; those require separate checks.
Your model library.
People. Access. Control.
Cost estimator · illustrative
Example: 1,000 input + 500 output tokens. Rates below are hypothetical.
Your live account and conversations are connected. Approve employee requests below.