Disclosures · updated 9 October 2026

Disclosures.

Every report we have sent, and where it stands.

Vendors hear first. Until the publication date a report shows only the vendor, the product and the dates; the detail is published on the date, or earlier if the vendor fixes it. The rules are in our disclosure policy.

ReportedVendor responsePublishedFixedResolved

Resolved means the vendor has shipped a fix and our re-test of the fixed release confirms it.

ID		Finding		Fix
`DSK-1`	DeepSeek DeepSeek Harness (dsh)	Sends your session, including tool results and the contents of files it reads, to DeepSeek by default; the telemetry switch does not stop it. Differs from DeepSeek's Data Processing Statement. Updated 26 Oct 2026. Read the note →	26 Sep 2026	Off switch added in web app, 28 Sep; upload still on by default	8 Oct 2026	Published
`XAI-1`	xAI Grok Build CLI	With telemetry on and a third-party model, sends that model API key to api.x.ai as the bearer; the binary carries upload code the published source does not. Read the note →	26 Sep 2026	No reply	26 Oct 2026	Published
`AID-1`	Aider maintainers Aider	At end of input (for example < /dev/null) answers yes to the Run shell command? prompt, so unattended runs execute model-proposed shell commands; --yes-always refuses them. Read the note →	26 Sep 2026	No reply	26 Oct 2026	Published
`ANT-1`	Anthropic Claude Code	Details held until the publication date.	26 Sep 2026		28 Oct 2026	Held
`QWN-1`	Qwen Qwen Code	Tells API-key users it collects nothing extra, but by default sends Alibaba Cloud a persistent installation ID, your model and your provider address; the telemetry switch does not stop it. Read the note →	27 Sep 2026	No reply; unchanged in 0.25.3 (19 Oct)	27 Oct 2026	Published
`CUR-1`	Cursor Cursor CLI	Documentation says terminal commands need your approval by default, but the first-start config allowlists Shell(ls); a Shell(python3) deny rule did not stop the agent running the interpreter another way. Read the note →	27 Sep 2026	Automated HackerOne notice only; no substantive reply	27 Oct 2026	Published
`CUR-2`	Cursor Cursor CLI	Details held until the publication date.	27 Sep 2026		26 Dec 2026	Held
`GEM-1`	Google Gemini CLI	Its documentation calls the usage statistics anonymized, but by default they carry a persistent installation ID and host hardware details, and the ID is sent with your API key on every model request. Read the note →	27 Sep 2026	Automated notice only; no substantive reply	27 Oct 2026	Published
`KIMI-1`	Moonshot AI Kimi Code	Details held until the publication date.	9 Oct 2026		8 Nov 2026	Held
`GOOSE-2`	Agentic AI Foundation (Block) goose	Details held until the publication date.	20 Oct 2026		19 Nov 2026	Held
`KIMI-2`	Moonshot AI Kimi Code	Details held until the publication date.	20 Oct 2026		19 Nov 2026	Held
`KIMI-3`	Moonshot AI Kimi Code	Details held until the publication date.	20 Oct 2026		19 Nov 2026	Held
`MINIMAX-1`	MiniMax MiniMax Code	Details held until the publication date.	20 Oct 2026		19 Nov 2026	Held
`AUG-1`	Augment Code Auggie CLI	Details held until the publication date.	24 Oct 2026		22 Jan 2027	Held

Dates are when we emailed the vendor's published security or privacy contact. Automated replies are listed as such. Vendors can write to research@agenticthinking.uk to respond, dispute a finding or ask for a re-test.

Disclosure policy

Disclosure policy.

Disclosure policy

Vendors hear first.

About 30 daysFor behaviour that differs from the vendor's documentation, privacy statement or settings, such as telemetry a privacy page says is not collected, or an opt-out that does not stop what it says.

90 daysFor security vulnerabilities: flaws that let someone other than the user read, change or take data, or run commands, beyond what the user allowed.

Time when it is neededWe extend the window when a fix is in progress, re-test fixed versions, and publish both results.

Nothing earlyUntil the date, affected results show only "Held" and the date, and no detail is published. The one exception: if users are at immediate risk or a flaw is being exploited, we may publish sooner, after telling the vendor.

We disclose by email to the vendor's published security or privacy contact. We do not resubmit through web forms or third-party platforms, including bug bounty platforms whose terms can restrict publication. An automated reply redirecting us elsewhere is treated as receipt and does not change our timeline.

Want an agent included, or want to dispute a result? Vendors can ask for a re-test after a fix.

Submit your harness Contact us

Or write to research@agenticthinking.uk.