Claude Fable security scan
Run a Claude Fable security scan with one command.
Claude Code supplies the Fable model and your login. Bex Security supplies the scan: repository scope, a sandbox that cannot write to your source, validation, and artifacts you can hand to another engineer.
Open source · Apache 2.0 · Report-only by default
npm install --global @bex-co/bex-security
claude auth login
bex-security scan . --agent claude --model fable --effort xhighModel and effort rows observed on September 2, 2026 from claude-agent-acp 0.70.0, bundled with @bex-co/bex-security 0.1.23-bex.1.
- Selected model
- Fable · claude-fable-5[1m]
- Agent runtime
- Claude Code · claude-agent-acp 0.70.0
- Project status
- Alpha
What actually runs
The session advertises its rows. Bex pins yours.
Bex asks the Claude ACP session which models and effort levels it offers, matches the values you pass by value or by name, and stops before the scan starts if a value is not on the list.
| Setting | Advertised rows | Selected by the command |
|---|---|---|
| Model
--model |
fable matches the row named Fable. The [1m] suffix is the 1M-token context window. | fable
claude-fable-5[1m] |
| Effort
--effort |
The session defaults to xhigh. Bex also accepts minimal, but this session does not offer it. | xhigh
xhigh |
A value the session does not advertise ends the run before any scan starts, and Bex prints the rows it did find.
Two owners, one boundary
Claude Code owns the model. Bex owns the scan.
Bex launches Claude Code as a subprocess with an explicit permission policy instead of trusting a prompt to keep your source untouched.
| Concern | Claude Code | Bex Security |
|---|---|---|
| Sign-in | Uses the login saved by claude auth login. | Starts the subprocess without Codex or provider API keys and stores no credentials of its own. |
| Model and effort | Discovers the rows your account can use. | Pins the values you pass and records the negotiated model and effort with the scan. |
| Repository files | May read the repository and run sandboxed commands. | Requires the sandbox, denies Edit and Write on the repository, and allows writes only in the scan output and state directories. |
| Findings | Investigates likely attack paths and drafts candidate findings. | Challenges every candidate, rejects claims without evidence, and seals the survivors as local artifacts. |
Quick setup
Three lines, in order.
Each line of the command has one owner. Install Bex, sign in to Claude Code once, then select the agent, the model row, and the effort level explicitly so the artifact records them.
- 01
Install Bex Security
Install @bex-co/bex-security from npm. The Claude ACP adapter ships inside the package; Claude Code itself must be installed locally.
npm install --global @bex-co/bex-security - 02
Sign in to Claude Code
Run claude auth login once. --agent claude delegates authentication and model discovery to that installation, and Bex stores nothing.
claude auth login - 03
Select the agent, model, and effort
Pass --agent claude --model fable --effort xhigh. Omit --model or --effort to accept the session's defaults; pass a row the session does not offer and the run stops before scanning.
bex-security scan . --agent claude --model fable --effort xhigh
Requires Node.js 22.13+ in the 22.x line, Node.js 24 or 26, and Python 3.10+. Scan only repositories you own or are authorized to assess.
Alpha integration. Claude Code sessions over claude-agent-acp are alpha: the vertical slice works and is being hardened with full-scan fixtures, but not every model and effort combination has passed a mature compatibility matrix.
Beyond the default scan
Flags that change what the scan does, and what it leaves behind.
Ordinary scans are report-only and print findings to the terminal. Each command below is an explicit decision; none of them is on by default.
| Command | What it decides |
|---|---|
| bex-security scan . --diff origin/main | Review only the changes since a base ref instead of the whole repository. |
| bex-security scan . --mode deep --max-time-hours 1.5 | Run a deep scan with a bounded discovery time; deep discovery stops after 96 hours unless you set a shorter limit. |
| bex-security scan . --json | Write structured results to stdout for CI and other tools. |
| bex-security scan . --fail-on-severity high | Exit with code 1 when a completed scan finds high or critical issues; incomplete scans exit with 2. |
| bex-security scan . --patch --patch-severity high | Fix high and critical findings. This is the first time repository files change, and only because you asked. |
| bex-security scan . --patch --create-pr | Commit the verified patch files and open a draft GitHub pull request. |
| bex-security export <results-dir> --export-format sarif | Export a scan's findings as SARIF, CSV, or JSON from its results directory. |
| bex-security scans compare BEFORE_ID AFTER_ID | Match findings by root cause across two scans and report which are new, persisting, reopened, or resolved. |
Frequently asked questions
Which Claude model does --model fable select?
The row the Claude ACP session names Fable, which was claude-fable-5[1m] on claude-agent-acp 0.70.0. Bex matches --model against the session's advertised rows by value or by name and refuses any other value, printing the rows it found.
Which credentials does a Claude Fable security scan use?
The login saved by Claude Code. Bex starts the Claude subprocess without Codex or provider API keys, and it stores no credentials of its own.
Can Claude Fable decline part of a security scan?
Yes. Claude Fable runs safety classifiers and may decline a request it judges harmful, even when you own the repository. Bex publishes a finding only after validation, so an investigation that stops early leaves no unverified claim behind.
Explore other AI security scans
Make Claude Fable reviews something another engineer can audit.
Install Bex Security when you want Fable's reasoning inside a bounded workflow that keeps your source read-only, validates findings, and leaves artifacts behind.
Star Bex SecurityOne workflow · Open agent layer · Validated evidence