Governance · Newsagent + Tumeryk
Every AI call checked. Every decision a person's. All of it on the record.
Newsrooms are now asked how they govern AI by their boards, their audiences, regulators and insurers. Newsagent answers with one gate in front of every model, approvals for anything that matters, and evidence for every step.
Watch
How the guardrails work
Eight short chapters: why newsrooms need guardrails, the single gate, what's checked, four checks in action, Newsagent + Tumeryk, trust scores, people deciding, and the evidence you get.
01AI now drafts versions, checks facts and style, summarises sources, and even reads the news.
What gets checked
Defaults shown beside each rule; each newsroom sets its own in Policies. Dashed rules need Tumeryk connected.
Before the model sees it
- Allowed tools, providers and modelsblockAI is only used for what the policy allows, with approved providers and models.
- Embargoed storiesflagA story still under embargo isn't sent to an outside model without the policy saying so.
- Confidential copyblockText marked confidential, not for publication, off the record or source protected.
- Hidden instructionsredactInstructions planted in wires, tips, transcripts and web pages, including invisible characters.
- Personal dataflagEmail addresses, phone numbers, tax file, Medicare and card numbers.
- Tumeryk input railsTumeryk decidesJailbreak, prompt injection and sensitive data, checked by Tumeryk's AI Trust Gateway.
- Model trust minimumflagCalls to a model whose Tumeryk AI Trust Score is below the newsroom's minimum.
Before anyone uses the answer
- Abuse and threatsflagProfanity the story didn't contain, and threats.
- Personal data it wasn't givenflagContact details or identifiers that weren't in what the model was sent.
- Echoed instructionsflagAn answer that repeats instructions aimed at a model.
- Chatbot answersflag“As an AI model…”: answered as a chatbot, not for the newsroom.
- Tumeryk output railsTumeryk decidesModeration and hallucination, checked by Tumeryk; a low hallucination score follows the low-trust rule.
What happens
When several checks fire, the strongest outcome wins.
- Allow1 of 4It goes ahead.
- Flag2 of 4It goes ahead, and it's marked for a person to review later. Flags never stop anyone working.
- Redact3 of 4It goes ahead with the risky part removed: personal data replaced with a label, planted instructions replaced with a notice.
- Block4 of 4It stops. Nothing is sent to the model or used, and the reason is shown.
Who decides
Agents and workflows propose; these people decide. One rejection ends it; admins can sign for any role.
| Request | Who signs off | Due |
|---|---|---|
| Publishing | Editor | 2 h |
| Agent proposal | Editor | 4 h |
| Workflow step | Editor | 4 h |
| AI-generated media | Editor | 24 h |
| AI policy exception | Admin | 24 h |
| Legal review | Editor and Admin | 8 h |
The defaults. Actions the policy pre-approves go ahead at once and are still recorded. Nothing AI writes is published without a person.
Evidence you get
What you can show an auditor, an insurer or the board.
- Verdict logEvery check: before or after the model, the rule, the outcome and who checked it (Newsagent or Tumeryk), with Tumeryk's reference.
- AI registerEvery AI run, agent step and workflow step, with its model, cost, checks, approvals and outcome.
- Audit logEvery change by people, workflows and AI, append-only. Blocks, flags and redactions are in it too.
- Your SIEMBlocks and flags forwarded to a security webhook as they happen.
- Framework mapControls mapped to OWASP's LLM Top 10, ISO/IEC 42001, the NIST AI RMF, the EU AI Act and Australia's Voluntary AI Safety Standard.
What Tumeryk adds
AI Trust GatewayEach exchange is checked by Tumeryk’s rails: jailbreak, prompt injection and sensitive data on the way in; moderation and hallucination on the way out. Its verdict (allowed, masked or blocked) and reference are kept with Newsagent’s own.
AI Trust ScoreA score for each model across nine dimensions, aligned with the NIST AI RMF, ISO 42001 and the EU AI Act, never more than a day old. Set a minimum to flag or block calls to models below it.
Runs where you need itTumeryk’s cloud, your private cloud or on premises. If it can’t be reached, your policy decides: carry on and flag it for review (the default), or stop. Newsagent’s own checks run on every call either way.
The gate
One gate for every AI call
Alpha Omega's operators, agents, workflows and AI presenter scripts all pass through the same checks, before the model sees anything and before anyone uses its answer. Each check allows, redacts, flags or blocks, and the strongest wins.
Before the model sees it
Is this tool and model allowed? Is the story under embargo? Is anything marked confidential? Personal data (including TFN and Medicare numbers), and instructions planted in wires, tips or transcripts.
Before anyone uses the answer
Abuse and threats, personal data the model wasn't given, echoed instructions and chatbot boilerplate.
With Tumeryk connected
Tumeryk's AI Trust Gateway runs its own input and output rails (jailbreak, prompt injection, sensitive data, moderation, hallucination) and returns a verdict and a trust score. Each newsroom chooses whether to carry on with a flag or stop if Tumeryk can't be reached.
People decide
Approvals in one queue
Agent and workflow proposals, AI-generated media, legal review and exceptions wait in one queue, for the roles your sign-off policy names. Each decision records who made it, when and why.
- Agents and workflows propose; nothing changes until it's approved
- AI presenter videos can't go to air until an editor approves them
- Pre-approve routine actions so people only see what matters
Evidence
A record you can show anyone
AI register
Every AI run, agent step, workflow step and AI-generated video: model, provider, cost, guardrail verdicts, approval and outcome.
Verdict log and audit trail
Every check is saved; anything not simply allowed is also in the audit log and can be forwarded to your SIEM.
Policies with history
AI use, sign-off and data rules, versioned, so you can show what applied on any day.
Incidents and legal holds
An AI incident register with severity and actions, and legal holds that stop records being deleted.
Frameworks
Mapped to the frameworks you're asked about
Each of Newsagent's controls is mapped to the items in these frameworks it helps you address. The mapping shows which control covers what; it isn't a certification.
| LLM01 | Prompt injection | Input guardrails · Tumeryk AI Trust Gateway |
| LLM02 | Sensitive information disclosure | Data rules · Output guardrails · Tumeryk AI Trust Gateway |
| LLM03 | Supply chain | Approved providers and models · Model trust scores |
| LLM05 | Improper output handling | Output guardrails · Human sign-off |
| LLM06 | Excessive agency | Agents propose, people decide · Human sign-off · Roles and permissions |
| LLM07 | System prompt leakage | Input guardrails · Output guardrails |
| LLM09 | Misinformation | Grounded, cited output · Human sign-off · Tumeryk AI Trust Gateway |
| LLM10 | Unbounded consumption | Spend limits |
| A.2 | Policies related to AI | Written policies |
| A.6 | AI system life cycle (operation and monitoring) | AI register · Input guardrails · Output guardrails |
| A.7 | Data for AI systems | Data rules · Retention |
| A.8 | Information for interested parties | AI-generated media labelled |
| A.10 | Third-party and customer relationships | Approved providers and models · Model trust scores · Encrypted provider keys |
| GOVERN | Policies, accountability and culture | Written policies · Human sign-off · Roles and permissions |
| MAP | Context and risks are known | Model trust scores · Approved providers and models |
| MEASURE | Risks are measured and tracked | Tumeryk AI Trust Gateway · Output guardrails · AI register |
| MANAGE | Risks are prioritised and acted on | AI incident register · Input guardrails · Spend limits |
| Art. 12 | Record-keeping | AI register · Audit log |
| Art. 14 | Human oversight | Human sign-off · Agents propose, people decide |
| Art. 50 | Transparency for AI-generated content | AI-generated media labelled |
| G1 | Accountability process | Written policies · Roles and permissions |
| G2 | Risk management | AI incident register · Model trust scores |
| G3 | Protect AI systems and data | Data rules · Encrypted provider keys · Newsroom isolation · Input guardrails |
| G4 | Test and monitor | Tumeryk AI Trust Gateway · Output guardrails · AI register |
| G5 | Human control | Human sign-off · Agents propose, people decide |
| G6 | Inform people about AI use | AI-generated media labelled |
| G8 | Supply chain transparency | Approved providers and models · Model trust scores |
| G9 | Keep records | AI register · Audit log · Retention |
Newsagent + Tumeryk
Why we built this with Tumeryk
Tumeryk’s AI Trust Score platform scores AI requests as they happen and returns a verdict, and scores models on nine dimensions aligned with NIST AI RMF, ISO 42001 and the EU AI Act. Newsagent plugs Tumeryk into its gate, so every AI call in the newsroom gets Tumeryk’s checks as well as Newsagent’s own, and shows each model’s trust score where you choose it.
- Newsagent's own checks always run, so nothing is unguarded if Tumeryk is off
- Tumeryk's verdicts and references are kept with Newsagent's record
- A minimum trust score per newsroom: flag or block calls to models below it
- Your choice when Tumeryk can't be reached: carry on with a flag, or stop
Talk to us about governing AI in your newsroom
We'll walk your team through Newsagent with your own desks, shows and channels, and talk through how you'd run it.