Public-Sector, Standards & Enterprise AI Safety Brief

Stop one risky AI action without shutting down the model.

Global model kill switches halt legitimate public and enterprise services alongside the risky call. Decern removes reusable credentials from the AI agent so the receiving system checks a single-use authorization per consequential action.

Prototype transparency: Decern is a frozen research prototype (258 workspace tests passing; 11 attack scenarios blocked pre-signing in internal testing, 0 unauthorized signing attempts on Bitcoin signet; U.S. patent application 64/130,485 pending). This general AI-agent interlock is an architectural adaptation proposed for contained sandbox evaluation rather than commercial procurement.

The Containment Gap

Why safety boundaries belong at the receiving system, not inside the agent.

State AI safety frameworks (including California Executive Order N-9-26 and GovOps's November 2026 blueprint, AI Safety in 2026 and Beyond) focus on loss-of-control containment, emergency shutdown controls, and licensed Independent Verification Organizations (IVOs).

01 · Reusable Credentials

Prompt filters fail when agents hold live keys.

If an AI agent holds a reusable API token or database credential, prompt injection or tool-loop drift can invoke consequential actions directly.

02 · Shutdown Hesitation

Operators hesitate to pull a global kill switch.

Shutting down a shared frontier model interrupts benign public benefits, infrastructure, or enterprise operations alongside the unapproved call.

03 · Mutable Host Logs

Agents can alter logs stored on their own host.

Supervisory records must be generated and timestamped outside the agent's reach so third-party evaluators can trust post-incident evidence.

Control Layer Enforcement Point Operational Effect Evaluator Evidence
Prompt Guardrails Inside the caller or agent wrapper. Bypassed when subverted code invokes reusable downstream keys directly. Mutable logs inside the same host environment.
Model-Level Kill Switch Model serving endpoint. Essential last-resort backstop, yet halts all benign traffic on that model. Coarse infrastructure uptime logs.
Decern Action Interlock Independent receiving gateway (agent holds 0 keys). Blocks specific unticketed, altered, or replayed actions while the model stays online. Signed allow/refusal receipts anchored externally for offline verification.

Interactive Containment Sandbox

Test 4 autonomous agent tool calls.

Click any scenario below to see how a receiving gateway checks a single-use authorization before executing a consequential action.

Receiving-Boundary Verdict Blocked: Parameter Swap
01 · Caller AI Agent 0 Live Keys
02 · Interlock Blocked Cold Parameter Mismatch
03 · Service Untouched 0 Executions
Agent Request:
Interlock Result:
Model stays online: routine workflows continue uninterrupted while the unapproved action halts.
For Evaluators & Engineers: Inspect Raw SHA-256 Receipt & Tamper Test +
APPROVED INTENT SHA-256:
RECONSTRUCTED ACTION SHA-256:

          

Independent Verification & Sandboxes

Built for third-party evaluators and narrow sandbox testing.

We are seeking technical critique and working-group input rather than commercial procurement.

Offline Receipts for Independent Verification Organizations (IVOs)

Because allow and refusal receipts are generated at the receiving boundary and anchored via outside RFC 3161 Timestamp Authorities (freetsa.org and ts.ssl.com), evaluators verify containment offline with zero calls to PriviNet.

Contained Non-Production Sandbox Proposal

Test one representative high-consequence boundary (data export, infrastructure config, or disbursement) against three clear metrics: 0 agent credentials, 0 unapproved executions, and 100% offline receipt verification.

Leadership: Founded by Brad Listermann (two decades of public-company governance and capital-markets leadership, including five years guiding iQSTEL to Nasdaq) alongside Co-Founder & COO Tal Segal and PriviNet's engineering team. Meet the team →

Direct: brad@privinet.net · (424) 250-0042

Request a 15-Minute Technical Exchange

For state AI safety teams, NIST and academic researchers, prospective IVOs, and enterprise CISOs.