[Use case] Using current model to audit and maintain an AI incident registry
Heat trend
Collecting trend data
The percentage is based on available heat signal, not comment count or independent people.
OpenAI model activity is surfacing — worth tracking for capability changes, ecosystem impact, and availability.
A use case for AI involves leveraging a current model to audit and maintain an AI incident registry. This process includes model-over-model quality assurance, where the current model reviews work from earlier models. It identifies unjustified confidence, preserves supported findings, incorporates new evidence, and updates the live research artifact to ensure accuracy and relevance.
https://preview.redd.it/wlsrx4l8n8nh1.png?width=1617&format=png&auto=webp&s=937c3646c60874e3d6495949bc002fd1490b8e1b
https://preview.redd.it/k5q1ud2on8nh1.png?width=1493&format=png&auto=webp&s=eafcbc24cf9fd43c94b5526c15db05628451eeaa
Revisited security research I'd previously conducted with earlier models across OpenAI, Google/Gemini and Anthropic.
Rather than starting the research again, I gave the current model the existing evidence and analysis and had it audit the previous work.
- challenged conclusions reached by earlier models and downgraded claims where the retained evidence didn't support them;
- reconstructed disclosure timelines and incorporated evidence that emerged after the original investigations;
- retrieved existing research from connected Notion pages and compared it against the retained record;
- maintained provenance distinctions between vendor statements, public reports, independently demonstrated findings and model inference;
I showed it screenshots of the live registry when the layout broke. It diagnosed the HTML/CSS problems, rewrote the affected components, added tabbed incident navigation, built dynamic status information and incorporated dated source links for continuing reports.
So within the same piece of work it moved between long-context reasoning, model-over-model QA, connected-app retrieval, web research, vision, evidence analysis, risk assessment, writing, coding and visual debugging.
The interesting part for me wasn't any individual feature. It was being able to use them together against the same persistent body of work without turning each stage into a separate workflow.
With all the discussion around Astra and its security capabilities, I am keen to see how:
https://preview.redd.it/6kvfd63wo8nh1.png?width=1356&format=png&auto=webp&s=ebabb555489d0e92172f36a9937d9f5446aa1a51
- Long-context reasoning: maintaining the OpenAI, Gemini and Anthropic cases simultaneously, comparing earlier conclusions with newer evidence, and keeping competing hypotheses separate.
- Critical analysis / self-audit: reviewing work produced by earlier models, identifying unjustified conclusions, and downgrading claims where the evidence did not support the original confidence level. This also included verification of previous bug-hunting work within code.
- Evidence classification: repeatedly separating observed fact → reasonable inference → hypothesis → unsupported claim / overreach.
- Risk assessment: reassessing technical security, privacy, governance, enterprise, systemic and disclosure risks as the available evidence changed.
- Temporal reasoning: reconstructing disclosure chronologies and evaluating later evidence without retroactively treating it as information available at the time of the original report.
- Cross-source synthesis: combining retained evidence, vendor responses, public reports, subsequent independent security research, and regulatory or assurance-framework material.
- Web search / browsing: locating and validating external evidence, including subsequent Google API-key research and continuing OpenAI billing/Codex reports.
- Connected apps / plugins: retrieving relevant material from connected workspaces and incorporating it directly into the analysis rather than requiring repeated manual transfer of source material.
- Notion retrieval: retrieving existing research and discussion material from Notion and auditing it against the retained evidentiary record.
- Vision: analysing screenshots of the incident registry, historical evidence, UI states and broken layouts, then incorporating visible details into technical and evidentiary analysis.
- Document / artefact interpretation: interpreting reports, timelines, disclosure correspondence, screenshots, tables and technical artefacts as structured evidence rather than treating everything as ordinary prose.