Watch the video, transcript and sources
Sunday, October 18, 2026 · The Daily Diff · Niko from Axrisi
https://www.youtube.com/watch?v=0vPlG6f14lg
- OpenCode Console reports a breach
+ Sphinx checks understanding before Git pushes
+ Pele renders Mermaid diagrams to themeable SVGOpenCode reports a security breach affecting 1.4% of Console users. Its edited notice identifies a subset with external providers linked and says those people received an email. The incident was still ongoing when we reviewed the statement. This is an explicitly requested follow-up to the overnight item in our earlier episode, where the overall verdict was NEEDS REVIEW.
That scope belongs in the headline. OpenCode's local coding agent and its optional hosted Console are distinct components. Installing the agent alone does not establish that you belong to the reported affected group. The notice gives a percentage of Console users without the total population, so an absolute victim count cannot be calculated from it.
The provider documentation explains the relevant trust boundary: Console stores a connected provider credential and serves requests through its gateway. Workspace members can use the models without seeing the upstream key. That's convenient administration. The gateway becomes a component whose handling of that credential matters.
The published statement supplies no technical reconstruction of how access was gained. It says the team is working with Anthropic, OpenAI and Google to remediate. A later report will be needed to assess the fix. We have no verified reimbursement promise or independently confirmed individual loss figure to add.
My practical advice is to review the official notification and provider activity records. If suspicious use appears, preserve the evidence, revoke the exposed credential through its provider and update affected integrations deliberately. This is general incident hygiene. It is not a newly published official customer remediation checklist.
A push can now come with an oral exam
Sphinx pauses a protected-branch push to ask whether you understand your agent's code. A separate reviewer reads the committed change without seeing the generating chat. Its questions concern behavior and failure cases. A wrong answer leads to another scenario; passing lets the same push continue.
The project's caveats are unusually useful. Installation changes the global Git hook path, so other repositories are initially included. The reviewer provider receives changed code and your answers. Filename exclusions cannot catch a secret pasted into an ordinary source file. Model calls have no default cap; a typical check uses five to eight requests plus roughly one per answer.
Local hooks can be bypassed. The examiner can be wrong. The documented sphinx yolo command logs a skip. We have not installed the tool or measured a production review improvement; its published worker/queue dialogue is explicitly illustrative. Your autocomplete has acquired a teaching assistant with its own disclaimer.
A smaller diagram dependency still needs an application policy
Obsidian's kepano released Pele, a Mermaid-to-SVG renderer whose colors can change through CSS without rendering the diagram again. We checked the real homepage editor: after changing the documented CSS variables, the same SVG node and its four labels remained in place. That demonstrates theme behavior. Render speed was outside this check.
The project's bundle comparison reports about 230 kB compressed against Mermaid's 1.6 MB. Its dramatic performance comparison remains vendor-reported; we did not reproduce it.
The security documentation explains what remains outside the renderer's filters. Displayed images can make network requests. The application owns supplied icon markup and internal-link navigation. There are options to disable links and images, plus source/output limits. A smaller dependency provides a useful contract while leaving real application work on the table.
Verdict: NEEDS REVIEW
I'd keep the Console incident under review until enough detail is available to assess the affected credentials and the fix. Identify which service actually holds your provider access before a dramatic headline chooses the threat model for you.
One final word deserves unpacking: distillation trains a model using another model's outputs. OpenCode associates the incident with a large-scale distillation attack. The label describes its alleged purpose; the public access path and full customer impact remain undisclosed.
And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.