Reachable demo based on Go testbed

Go Testbed Expected Results

A compact demo for CI scanning and agent-driven remediation. It proves that Reachable can identify the release-blocking queue from database evidence, generate a bounded remediation task, let an agent patch a review branch, and rescan that branch to zero release blockers. Workflow security is represented here only by a compact customer-facing subset, not the full support corpus or remediation experiment matrix.

Evidence Links

This page explains the demo contract. The current run status, source code, workflow definition, expected contract, and published proof artifacts are linked here so the evidence can be traced.

Workflow runs Workflow code Repository code Branded contract page Machine-readable contract Last published proof page Published artifacts DB verdict artifact

DB Signals

Expected raw scanner truth on main.

Release Blockers

Release-blocking queue before remediation.

Evidence Rows

DB evidence rows used in the public proof.

Proof Target

Release blockers after remediation.

What Had To Be Fixed

This is the release decision view. These are the baseline problems that must disappear from the remediation proof scan before the branch is safe to review and merge.

Release category	Rows before	Expected proof	Remediated?	Meaning
Exploitable blockers	3	0	Yes	Enzo attacker proved external control and impact. These are mandatory fixes.
Not-attacked / exposure blockers	9	0	Yes	No attacker proof is required for reachable CVEs, embedded secrets, PII/DLP exposure, or AI-boundary authority risk. They still block release while present.
Defended attacker evidence	9	0 blockers	No, unless a fix naturally removes them	Reachable found the path, but Enzo attacker could not prove a practical attack. These rows are audit evidence, not release blockers.
Filtered fixture evidence	7	Nonblocking	No	Synthetic not-reachable or non-production markers prove scanner coverage and are not part of the fix queue.
Workflow-security guidance	25	Nonblocking admin guidance	No autonomous code fix	The demo keeps a compact CI/CD workflow-authority subset visible for review without turning this repo into the workflow-security lab.
Total raw DB signals	53	0 release blockers	12 remediated blockers	The proof scan passes when the release-blocking queue is zero.

Fix category	Rows	Exploitability	Why it blocks release	Expected proof after remediation
Application attack paths	9	Exploitable	Enzo attacker evidence says reachable CWE/AI paths are attackable, including command execution, SSRF/fetch, and externally visible error disclosure.	Those rows are absent from the proof scan.
Dependency risk	1	Reachable, not attack-proofed	A reachable Go dependency has a fix-available vulnerability.	The fixed dependency version is used and the CVE row is absent.
Secret exposure	1	Exposure	A synthetic production-shaped token is embedded in application code.	The token row is absent; fixture-only markers remain non-actionable.
DLP / PII exposure	2	Exposure	Synthetic personal data is logged or sent to an outbound endpoint.	Raw personal data exposure rows are absent.
AI boundary and tool-risk rows	5	Exploitable / exposure / authority risk	LLM calls, sensitive context, tool instructions, and unguarded user flows cross AI trust boundaries.	The AI rows are absent or covered by the underlying code fix.
Total release blockers	12	Actionable	Baseline queue that must be fixed before release.	0 release blockers after remediation.

Defendable And Nonblocking Evidence

These rows are still part of the baseline evidence, but they are not the final release-blocking queue. They explain why the raw database has more signals than the fix queue.

Evidence category	Rows	Meaning
Defended attacker rows	9	Reachable paths reviewed by Enzo attacker and classified as defended, not exploitable.
Non-production or not-reachable fixture markers	7	Synthetic detector coverage rows such as workflow variable names and fixture markers.
Workflow-security guidance subset	25	A compact CI/CD authority-review subset kept in the public demo contract, without importing the larger support corpus.
Total nonblocking evidence	41	Tracked for auditability, not counted as release blockers.

Why Defended Rows Do Not Block Release

Reachable keeps defended rows in the evidence when it finds a security-shaped sink, but Enzo attacker cannot prove an external attacker can drive that sink to impact. That gives the release owner an audit trail without forcing a non-exploitable item into the fix queue.

Defended row	Why it was found	What blocks the attack
`CWE/200` `internal/handlers/cve.go:16`	The handler returns an internal request-body read error string.	The body is read through `io.LimitReader`, so the request body is bounded; this path is an I/O failure path, not attacker-controlled parser content.
`CWE/200` `internal/handlers/cve.go:38`	`language.Parse` can produce detailed parsing errors.	The route first calls `safety.AllowedLanguageTag`, which allows only `en`, `en-US`, `fr`, `fr-FR`, `es`, and `es-ES`. Attacker-chosen language tags do not reach the parser.
`CWE/200` `internal/handlers/suspicious.go:24`	The tool-fetch route returns an `os.Create` error.	The local file path is fixed with `filepath.Join(os.TempDir(), "reach-testbed-tool.bin")`. The caller controls the source URL, not the local file path that could shape this error.

Decision rule: if Enzo attacker proves external control and impact, the row is exploitable and blocks release. If a guard, fixed value, or internal-only branch prevents that proof, the row stays visible as defended evidence.

Baseline Signal Breakdown

Family	Expected signals	Release blockers	Exploitability dimension	Nonblocking evidence	Expected proof
CVE	1	1	Reachable, not attack-proofed	0	Fixed dependency version; CVE row absent from proof scan.
CWE	15	3	3 exploitable, 9 defended, 3 filtered	12 nonblocking rows	Command execution, SSRF/network fetch, and blocking error-disclosure rows absent from proof scan.
Secret	3	1	Exposure	2 filtered rows	Synthetic production token absent; workflow-template noise is not part of the baseline anymore.
DLP	2	2	Exposure	0	Synthetic personal data exposure rows absent.
AI	7	5	Reachable authority/exposure plus unguarded flow	2 nonblocking rows	AI-boundary and unguarded-flow rows absent or covered by underlying code fixes.
Workflow Security	25	0	Workflow authority review	25 guidance rows	Retained as a compact customer-demo workflow-security subset; not part of the autonomous code-remediation queue.
Total	53	12	3 exploitable rows plus exposure/reachable blockers	41	0 release blockers after remediation.

Trust model: the expected results are declared in expected/baseline.json. CI validates the vulnerable baseline scan database against that contract, then validates the remediation proof scan database to confirm the same release blockers are gone. The proof is a database comparison, not a hand-written status claim.

Expected Findings

The table separates the dimensions a release owner needs to understand: raw DB signal count, risk, reachability, and exploitability. A row with more than one DB signal is an explicit group, not a hidden assumption.

ID	DB signals	Type	Risk	Reachability	Exploitability	Location	Explanation	Remediation
GO-CVE-01	1	CVE-2022-32149 / `golang.org/x/text`	High	Reachable	Not attack-proofed	`internal/handlers/cve.go`; `go.mod`	A public language parsing route exercises an old dependency with a denial-of-service advisory.	Upgrade `golang.org/x/text` to a fixed version and keep route-level input validation.
GO-CWE-01 Deferred	1	CWE / command injection	Critical	Reachable	Exploitable	`internal/handlers/cwe.go`	A request parameter is concatenated into a shell command. A caller could turn a diagnostic endpoint into command execution. This case remains in the vulnerable baseline contract, but is temporarily deferred from the autonomous-remediation proof because the current agent loop leaves `DiagnosticPing` unchanged when `SafeDiagnosticPing` already exists.	Manual-review/deferred for this demo. The reach-core follow-up is to tighten the remediation prompt so the agent modifies `DiagnosticPing` itself.
GO-CWE-02	1	CWE / user-controlled URL fetch	Critical	Reachable	Exploitable	`internal/handlers/suspicious.go`	An admin route downloads from a caller-supplied URL. This models unsafe tool staging and SSRF-style fetch behavior.	Restrict sources to a trusted allowlist, require authentication, verify checksums/signatures, and avoid arbitrary outbound fetches.
GO-CWE-03	1	CWE / SSRF HTTP client	Medium	Reachable	Exploitable	`internal/handlers/suspicious.go`	User input reaches an HTTP client, so server-side infrastructure could be asked to call untrusted destinations.	Use URL validation, deny private/internal ranges, enforce trusted schemes/hosts, and add timeouts.
GO-CWE-04	3	CWE / error disclosure	Medium	Mixed blocking/nonblocking	Mixed exploitable/defended	`internal/handlers/cve.go`	Parser errors are returned directly to clients, potentially exposing implementation details; non-attacker-controlled instances are retained as nonblocking evidence.	Return generic client errors and log details internally.
GO-CWE-05	3	CWE / error disclosure	Medium	Reachable	Exploitable	`internal/handlers/ai.go`	JSON decoding errors are returned directly from AI endpoints.	Return generic bad-request text and preserve details only in structured logs.
GO-CWE-06	3	CWE / error disclosure	Medium	Mixed blocking/nonblocking	Mixed exploitable/defended	`internal/handlers/suspicious.go`	Network, file, and copy errors from the tool-fetch path are exposed to callers; one internal-only instance is retained as nonblocking evidence.	Return generic operational errors; keep internal details in logs or audit events.
GO-SECRET-01	1	Secret / GitHub token shape	Medium	Reachable	Exposure	`internal/handlers/secrets.go`	A synthetic GitHub-shaped token is embedded in code and returned by an API. In a real system this would be a credential leak.	Rotate the value, remove it from code, load it from a secret manager, and never return it in responses.
GO-SECRET-02	1	Secret / AWS access key shape	Info	Not reachable	Not applicable	`internal/handlers/secrets.go`	An AWS-shaped synthetic marker is present for detector coverage but is filtered as non-actionable in the latest proof.	Keep only synthetic test markers in fixtures; never put real cloud credentials in source.
GO-SECRET-03	1	Secret / static token detector overlap	Info	Not reachable	Not applicable	`internal/handlers/secrets.go`	A generic token-shape detector overlaps the synthetic GitHub PAT fixture on the same line. The specific GitHub PAT row is the one that remains actionable.	No separate code fix required. This overlap remains audit evidence only.
GO-DLP-01	1	DLP / PII to log	Critical	Reachable	Exposure	`internal/handlers/dlp.go`	Synthetic SSN and date-of-birth values are written to logs. In production this would create regulated-data exposure.	Mask sensitive values, minimize logging, and add structured audit logging without raw identifiers.
GO-DLP-02	1	DLP / PII to outbound HTTP	Critical	Reachable	Exposure	`internal/handlers/dlp.go`	Synthetic personal data is sent to an external analytics endpoint.	Remove raw PII from outbound telemetry, tokenize fields, and enforce data-sharing controls.
GO-AI-01	1	AI / LLM API call with sensitive context	Critical	Reachable	Exposure	`internal/handlers/ai.go`	User-controlled prompt content is sent to an LLM call in an admin-style context.	Separate system and user messages, treat user content as data, and apply policy checks before model calls.
GO-AI-02	1	AI / agent tool instruction risk	Critical	Reachable	Reachable authority risk	`internal/handlers/ai.go`	User input is mixed into an internal automation-agent tool specification.	Use constrained tool schemas, allowlisted actions, explicit authorization, and policy checks.
GO-AI-03	1	AI / unguarded flow to command execution	Medium	Reachable	Exploitable	`internal/handlers/cwe.go`	Reachable taint flow confirms the command-injection path has user-controlled input.	Fixed by the command-injection remediation.
GO-AI-04	1	AI / unguarded flow to error/output response	Medium	Reachable	Exploitable	`internal/handlers/cwe.go`	User-controlled diagnostic behavior can influence returned output.	Fixed by removing shell execution and normalizing errors.
GO-AI-05	1	AI / unguarded flow to network fetch	Medium	Reachable	Exploitable	`internal/handlers/suspicious.go`	User input controls the outbound fetch destination.	Fixed by the URL allowlist and SSRF controls.
GO-WORKFLOW-01	25	Workflow Security / CI/CD authority paths	Mixed	Reachable workflow authority	Admin guidance	`.github/workflows/reachable-workflow-security-fixture.yml`	The compact workflow-security fixture proves command injection, code injection, broken authorization, cross-workflow privilege escalation, untrusted checkout, OIDC/secret authority exposure, action pinning drift, and artifact-poisoning companion findings. Workflow files are owned by the dedicated `workflow_security` lane rather than the generic secret lane.	Review workflow trust boundaries, protect low-trust triggers, pin actions, narrow secrets/OIDC scope, and keep privileged sinks behind manual or protected release boundaries.
Total	53	Raw DB signals represented by the grouped expected-finding contract.

This table is the expected-finding contract: 18 grouped rows representing 53 raw DB signals. Workflow security remains a compact representative subset here, not the full support corpus. The branded contract page is this page; the source Markdown is stored in the repository, and the machine-readable contract is expected/baseline.json. CI checks the scan database against that contract before and after remediation.

CI Flow Overview

The flow below is supporting context. The verdict itself comes from the scan database and the expected-finding contract above.

Reachable CI autoremediation process flow

Workflow Details

This page is the contract and evidence view. The CI mechanics, inputs, guardrails, and repository layout live in the repository README and the Reachable Remediation Template workflow.