AI Intelligence

Claude Approvals Exposed Passwords

7 stories · ~7 min read

Claude Approvals Exposed Passwords

Listen

If You Only Read One Thing

The permission prompt can create the leak it's supposed to prevent. Apiiro's investigation found credentials saved inside Claude Code's remembered approvals, then published to GitHub. Open WebUI's workstation agents bring the same question into browser chat: what else happens when an action is allowed? The consequential change is how approval reaches files and machines, rather than how convincingly an assistant asks.

Claude Code Saved Secrets Alongside Permission

Claude Code can turn a one-time database password into a lasting permission rule. The model doesn't need to disobey anyone. A developer approves a command containing a secret, and the runtime saves the command so it won't ask again. Anthropic's permission documentation describes permanent command approvals; a later commit can publish their contents.

Apiiro's November 5 research, updated November 8 with Anthropic's response, found secrets in 1,122 of 104,185 public .claude/settings.local.json files. Those are exposed files, not 1,122 confirmed intrusions. The population is public GitHub copies of this particular file, not all Claude Code users.

The reassuring part of the interface hides a second decision. “Run this command” authorizes execution; remembering the exact command also authorizes storage of its contents. The password moves from a credential source into a project artifact. From there, ordinary collaboration can distribute it without any further model decision about disclosure.

Claude Code does try to exclude its local settings through the developer's global Git ignore configuration. Apiiro describes cases that protection misses, including already-tracked files. A machine-specific exclusion also doesn't travel with the repository. Calling a file “local” tells a person what it was intended for; Git follows its tracking rules.

Adding a scanner doesn't fully repair that mistake. In Apiiro's comparison, both Gitleaks and TruffleHog missed 215 of the 1,122 affected files. Passwords embedded in shell commands don't necessarily resemble recognizable provider keys. A clean scan can therefore coexist with a real database credential in an approved command.

That gives the practice a precise shape. A command recorded as psql "$DB_URL" can retain a reference rather than the connection string itself. Keeping local approval files out of tracking prevents another copy from entering the repository. Neither measure removes a credential already present in history, and variable references don't stop a process from printing secrets later. The improvement is narrower: approval stops manufacturing another plaintext credential store.

Boris Cherny's response points to automatic Git exclusion and attributes the exposed copies to manual creation and check-in. Apiiro disputes that explanation. Cherny also discussed possible scanning and auto-mode changes; those aren't a shipped cure. A decisive repair would prevent a remembered approval from retaining the literal secret, including when the settings file is already tracked.

Open WebUI Puts Workstation Agents in Browser Chat

Open WebUI now lets a browser conversation drive coding agents on a real computer. Its November 8 release makes Claude Code, Codex and Pi selectable through a connected Open Terminal. The change takes remote agent work beyond watching a terminal: tool approvals and voice interaction sit in the chat interface.

The 0.13.1 release also introduces Tether, a desktop companion connecting a workstation to Open WebUI. Tether connects outward, avoiding an inbound port-opening step. Its published desktop downloads cover macOS, Windows and Linux. This is available software, although I haven't run an end-to-end session.

The consequential choice is where the work happens. A hosted agent usually starts with a provisioned workspace. A workstation agent can use the repository, files and installed tools already on that machine. Opening the conversation from a phone no longer means copying the project into a phone-friendly environment. It means reaching the computer that already has it.

That convenience changes what the chat account controls. A request that previously produced text can now cause a process to edit real files. An outbound connection simplifies networking; it doesn't establish which local resources the agent can touch. Claude's approval problem concerns what gets saved. Open WebUI raises the next question: which computer can act on an approval?

The interface also makes two different services look similar. If the terminal disappears, agent conversations fall back to the default model while earlier replies retain the agent's name. The merged fallback fix describes a saved agent default surviving after its terminal was disabled. A reply can continue even though the computer capable of doing the work is gone. Conversation continuity is weaker evidence than execution continuity.

There are costs beyond model tokens. Tether's license permits free personal, noncommercial use, but commercial or organizational production use requires an Enterprise License. The same day's 0.13.2 patch fixes chats failing to load over plain HTTP at LAN addresses. Availability doesn't yet establish dependable operation across network conditions.

The adoption case is strongest where work depends on an existing machine's setup. A disposable task with a clean bootstrap has less reason to inherit that workstation. The revealing failure test is disconnecting the terminal during an assignment: subsequent replies need to make clear whether any agent can still execute the requested work.

The Contrarian Take

Everyone says: Keeping a person at the approval prompt prevents an agent from leaking credentials.

Here's why that's wrong: Apiiro's failure happens through an approved action. The human can correctly authorize a database check without realizing that the runtime will preserve its password in a file. Reading the command more carefully might catch the secret, but it doesn't make the interface explain that second effect. The safeguard has to cover what approval records and distributes, as well as what it executes. More interruptions won't fix a storage decision the prompt never makes explicit.

Under the Radar

  • Arete changes Qwen's answers without demonstrating a general coding upgrade. Nous Research's November 6 Arete-27B release reports state-framed answers falling from 22% to 0.9% on its own 803-prompt evaluation. The paired models used matching decoding settings; the lab also designed the rubric. Downloadable weights and evaluation data make the claim inspectable. This is a candidate for knowledge work involving those subjects, with no independent evidence of a broader capability gain.

  • AWS keeps failed remote work visibly unfinished. Its November 6 AgentCore coding-agent sample records the verification command, exit code and attempts in a pull request. The PR stays a draft unless verification passes; the agent never merges. The refinement is that remote execution produces a reviewable account of failure too. A passing command still proves only the behavior that command checks, and deploying the sample incurs AWS charges.

Quick Takes

  • DeepSeek's Harness is becoming a plugin platform. In its November 8 community update, DeepSeek reports 1.54 million daily users through its official API as of November 4, with about 35,000 third-party plugins and 60% of active users having built or installed one. Those are company figures, not audited productivity results. A planned marketplace would make finding and maintaining plugins part of the runtime's appeal; no launch date is supplied. (Source)

  • Microsoft sees ARTEX reaching financial-sector victims. Its November 6 account identifies Cambodia-based Storm-3358 using the framework with conventional tools to exploit internet-facing systems and install remote-access software. This adds a named criminal deployment to October 31's coordination evidence. It doesn't measure how much AI accelerated the campaign. Agent security now includes the familiar persistence tools an agent can install, not just its own continued operation. (Source)

  • Small decision models buy speed at a measurable accuracy cost. Artificial Analysis's November 8 comparison puts Drex Max v1.1 at 73 and GPT-6 Luna at 76 on its confidence-aware Decision Index. Luna costs $1,530 per million tasks versus $102 and takes 11.34 seconds versus 275 milliseconds. But Drex passes only 63% of documents requiring all 15 answers correct; Luna passes 86%. Cheap routing and complete document processing justify different model choices. (Source)

The Thread

Agent interfaces increasingly hide decisions that used to require separate actions. Remembering an approval writes a file. Selecting an agent reaches a computer. Continuing a conversation may silently change which system answers. That makes the transition itself worth testing: what changed when the user clicked, and what survives when a dependency disappears? The further consequence is organizational. A developer may approve a command, while somebody else owns repository publication, device access or account recovery. Each person can do their assigned job correctly and still leave an unsafe combination. Testing those handoffs becomes necessary because the interface has bundled several owners' decisions into one apparently simple action.

Prediction Ledger

Weekly Scorecard

  • Decisions API would leave beta by November 7. Made October 7, medium confidence. Correct: OpenAI's October 22 changelog records general availability. The forecast followed a provider-owned milestone that subsequently shipped.
  • Exa would release ATLAS tasks, answer tables and a runnable grader by November 9. Made October 9, medium confidence. Pending verification: the primary announcement still promises those artifacts; this review didn't establish their release. The deadline is today, so the call cannot yet be scored wrong.

Other overdue forecasts remain pending where dated evidence doesn't establish the outcome. A current product page cannot, by itself, prove what was available at an earlier deadline. No new forecast is added to that backlog today.


Issue date: 2026-11-09 · Generated: 03:45 AM ET.

Tomorrow morning in your inbox.

Subscribe for free. 10-minute read, every weekday.