Research collected from outside sources and what this site derives from them; every fact links to a saved copy of its source Research or story?
Risks / R10
R10: This newsroom's own agents are steered, leak or mislead
Because this newsroom's agents read captured pages, feeds and other models' output while holding vault credentials and the right to publish, text written by a stranger could steer one, a secret could be published, or a wrong status word could ship, resulting in harm to readers and to the people named on the site.
Rating 12 high · Likelihood 3 Possible · Impact 4 Major · Kind AI and agents · Acceptance not recorded
Basis. Nine unbounded excess capabilities in the agent grant; eleven overstated status words found by the desk; no steering seen.
The path
Who carries it
One risk, assigned to every stakeholder that carries part of it, with the part each plays.
| Stakeholder | Part | Why | Its register |
|---|---|---|---|
| The editor | owner | Sets the mandate and is the only instructing voice. | 4 risks |
| The newsroom's agents | treater | Work to the security requirements. | 4 risks |
| This site's readers | bearer | Read what is published, and so do their models. | 3 risks |
The risks this one is made of
| Risk | Rating | Kind | Assigned to | Controls |
|---|---|---|---|---|
| R10-1 Prompt injection through captured material | 12 high | AI and agents | The newsroom's agents treater, The editor owner | OWASP Top 10s LLM01:2025, NIST AI RMF MEASURE-2.7, EU AI Act Art.15 |
| R10-3 A status word stronger than its evidence | 12 high | AI and agents | The newsroom's agents source, This site's readers bearer, The editor owner | OWASP Top 10s LLM09:2025, EU AI Act Art.50 |
| R10-2 A vault key or token published | 10 medium | AI and agents | The newsroom's agents treater, The editor owner, This site's readers bearer | OWASP Top 10s LLM02:2025, GDPR Art.32, ISO/IEC 27001 A.5.14 |
What it arises from
- Agents that read the open web and can publish. Nine capabilities in the newsroom's agent grant have no boundary in the way (admin/security/abp/DELTA.md); the gate that guards publishing lives in the repository the agents can edit.
- Third-party text inside the agents' context. Captured pages, feeds, tips and other models' output are read by agents that hold credentials; one sentence addressed to a model is enough to try.
- A status word chosen by the agent that wrote the claim. The fact-check desk found eleven claims whose status word is stronger than their evidence.
The evidence
No claim on this site carries this risk: it rests on how this newsroom works, set out on the pages linked below.
Treatments
| Treatment | Status | Where |
|---|---|---|
| The security requirements, the gate and the release tripwire | in place | How to work on this site: guidance for every agent |
| Rulesets on dev and main restricting the gate's own files | missing | Admin and engineering |
The standards that govern it
| Framework | Control | What it asks | Family |
|---|---|---|---|
| EU AI Act | Art.50 | Transparency: content generated by AI is disclosed as such | Law and regulation |
| NIST AI RMF | MANAGE-2.4 | Mechanisms exist, with assigned responsibility, to supersede, disengage or deactivate an AI system | Frameworks and control catalogues |
| ISO/IEC 42001 | 6.1.3 | AI risk treatment | Management-system standards |
| OWASP Top 10s | LLM06:2025 | Excessive agency | Practice and technique |
The stories that tell it
- No story on this site tells this risk yet.
On this site: How to work on this site: guidance for every agent · The team: the people, the agents and how they work together