Back to skill

Security audit

sn-ppt-workbench

Security checks for vulnerabilities and agentic risk

Overview

This presentation editor needs review because reachable users can access sensitive controls and files, while its agent integrations grant excessive authority.

Prefer a release that fixes authentication, deck-path confinement, automatic permission approval, conversation sharing, and profile replacement. Do not expose this version through LAN access or public forwarding. Any evaluation should use an isolated environment without personal credentials or private conversations, with deck and profile backups. Loopback binding reduces exposure but does not repair these flaws. The partial scanner report was supplemented with direct code inspection and an isolated path check; external agents and full browser behavior were not exercised. The evidence supports Review, without establishing malicious intent.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (2)

T09 · Insecure Skill Coding Practices

Error
Location
workbench-runtime/dist-server/index.mjs:39230
Finding

Sensitive HTTP APIs lack client authentication

Content
View full analysis

Impact

Unauthorized disclosure and modification of presentation content, server shutdown, and submission of agent requests using the owner's configured bridge. Agent-side effects depend on the configured provider and its own restrictions.

T09 · Insecure Skill Coding Practices

Error
Location
workbench-runtime/dist-server/index.mjs:33089
Finding

Unvalidated deck IDs bypass file-serving directory confinement

Content
View full analysis

Impact

Disclosure of arbitrary files readable by the server account. For example, on POSIX, sufficient decoded parent components can select the filesystem root and a relative file path such as etc/passwd can then pass the existing containment check. Sensitive configuration and credentials are exposed if readable and their paths are known.

Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
Findings (39)

Tainted flow: 'request' from os.environ.get (line 265, credential/environment) → urllib.request.urlopen (network output)

Critical
Category
Data Flow
Confidence
98% confidence
Finding

The helper sends a resolved gateway credential to a configurable URL before establishing that the destination is trusted; checking the response for 'hermes-agent' happens after disclosure. A manipulated gateway URL can therefore receive credentials loaded from Hermes configuration, and non-loopback HTTP also exposes them in transit.

Content

Scanner excerpt · scripts/open_workbench.py (reported line 271)May include surrounding context.

python
method="GET",
    )
    try:
        with urllib.request.urlopen(request, timeout=4) as response:
            body = response.read(10000).decode("utf-8", errors="replace")
            return response.status == 200 and "hermes-agent" in body
    except (OSError, urllib.error.URLError, urllib.error.HTTPError):

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The supplied chunk tests and troubleshoots a local agent gateway rather than opening or editing a presentation. It does not reference a deck directory, locate the latest deck, or interact with PPT Workbench presentation content. Although gateway diagnostics could support a broader application, this chunk provides no connection establishing these operations as implementation details of the declared presentation workflow.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

Most visible functionality supports the declared presentation workflow, including file previews, editing-related chat, and deck auditing. However, pi() explicitly sends a server-shutdown request, a process-control capability not described by opening, inspecting, or editing a deck. gi() also submits generation-style selections rather than edits to an existing presentation. The supplied chunk establishes these client-side request capabilities; it does not show their backend implementations.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

Deck discovery, runtime validation, and launching the editor support the declared purpose. However, automatically accessing external credential stores and using those credentials to discover and configure an agent gateway extends beyond opening or visually editing an existing deck. This happens without a separate request to enable agent integration; credential loading occurs even before argument parsing and dry-run handling. The supplied code does not establish malicious exfiltration or show what the downstream agent executes, but the credential access and authenticated integration are concrete, security-relevant capabilities absent from the description.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

Most visible functionality supports the declared purpose. However, component J includes runtime status reporting and an explicit stop-server confirmation workflow. Its confirmation handler calls pi() and subsequently navigates to about:blank. Although pi() is defined outside this chunk, the surrounding controls clearly expose server shutdown as an intended capability. Runtime shutdown is materially distinct from opening, inspecting, or editing a presentation and is not disclosed in the description.

Content

No source excerpt is available for this finding.

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 119)May include surrounding context.

md
When bypassing `open_workbench.py` and invoking the workbench launcher directly, use subcommands:

Credential Access

High
Category
Privilege Escalation
Confidence
92% confidence
Finding

This path is used by the unconditional hydration routine to read Hermes secrets from HERMES_HOME when merely opening a deck. It contributes to unnecessary credential access and subsequent bridge use, although the path expression alone is not evidence of malicious credential theft.

Content

Scanner excerpt · scripts/open_workbench.py (reported line 205)May include surrounding context.

python
"""Yield Hermes env files that may contain the Gateway bridge secret."""
    value = _clean_env("HERMES_HOME")
    if value:
        yield Path(value).expanduser() / ".env"
    if os.name == "nt":
        local_app_data = os.environ.get("LOCALAPPDATA", "").strip()
        if local_app_data:

Credential Access

High
Category
Privilege Escalation
Confidence
92% confidence
Finding

On Windows, the default opening flow searches the user's Hermes application configuration and imports gateway credentials. In this presentation-opening context, that access unnecessarily increases the authority available to the runtime and its agent bridge.

Content

Scanner excerpt · scripts/open_workbench.py (reported line 209)May include surrounding context.

python
if os.name == "nt":
        local_app_data = os.environ.get("LOCALAPPDATA", "").strip()
        if local_app_data:
            yield Path(local_app_data) / "hermes" / ".env"
    yield Path.home() / ".hermes" / ".env"

Credential Access

High
Category
Privilege Escalation
Confidence
94% confidence
Finding

The helper automatically reads the user's ~/.hermes/.env as a credential source even though local presentation opening does not require gateway access. The imported secrets can then be used in authenticated probes and passed to the Workbench runtime, extending the consequences of an endpoint or runtime compromise.

Content

Scanner excerpt · scripts/open_workbench.py (reported line 210)May include surrounding context.

python
local_app_data = os.environ.get("LOCALAPPDATA", "").strip()
        if local_app_data:
            yield Path(local_app_data) / "hermes" / ".env"
    yield Path.home() / ".hermes" / ".env"


def _load_env_file(file_path: Path) -> dict[str, str]:

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · workbench-runtime/dist-server/index.mjs (reported line 28336)May include surrounding context.

js
exports.compileToken = compileToken;
    function compileRules(rules, options, context) {
      var _a;
      return rules.reduce(function(previous, rule) {
        return previous === boolbase_1.default.falseFunc ? boolbase_1.default.falseFunc : (0, general_js_1.compileGeneralSelector)(previous, rule, options, context, compileToken);
      }, (_a = options.rootFunc) !== null && _a !== void 0 ? _a : boolbase_1.default.trueFunc);
    }

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · workbench-runtime/dist-server/index.mjs (reported line 34258)May include surrounding context.

js
"gemini-3.1-pro-preview",
  "glm-4.6v"
];
var envLocalPath = path3.resolve(process.cwd(), ".env.local");
var validBackends = /* @__PURE__ */ new Set(["hermes", "direct", "pi-agent", "gateway"]);
function parseEnvFile(filePath) {
  if (!existsSync3(filePath)) {

The companion bridge automatically approves agent permission requests instead of restricting capabilities to presentation editing.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

The ACP client selects an allow option for permission requests without obtaining user approval or checking the requested action. A compromised or prompt-injected agent could therefore receive permission for destructive operations or sensitive access beyond presentation editing; the separate workspace check on client-mediated file operations does not constrain every capability an adapter might request.

Content

No source excerpt is available for this finding.

The workbench discovers and reads other agent conversations and forwards their excerpts into companion requests.

High
Category
Excessive Agency
Confidence
97% confidence
Finding

The workbench discovers conversations using loose title and workspace substring matches, reads their messages, and automatically appends excerpts to companion requests. Opening a presentation does not establish authorization to disclose other conversations, and ambiguous matches can expose unrelated sensitive content or introduce attacker-controlled instructions into an agent with editing capabilities.

Content

No source excerpt is available for this finding.

Hidden Instructions

High
Category
Prompt Injection
Confidence
85% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · workbench-runtime/dist/assets/MarkdownArtifactViewer-DxLTmGwT.js (reported line 2)May include surrounding context.

js
on;払;䑍ot;䄗ĀDrṁṅot;扒;쀀𝔢ƀ;rsṐṑṗ檚ave耻è䃨Ā;dṜṝ檖ot;檘Ȁ;ilsṪṫṲṴ檙nters;揧;愓Ā;dṹṺ檕ot;檗ƀapsẅẉẗcr;䄓tyƀ;svẒẓẕ戅et»ẓpĀ1;ẝẤijạả;怄;怅怃ĀgsẪẬ;䅋p;怂ĀgpẴẸon;䄙f;쀀𝕖ƀalsỄỎỒrĀ;sỊị拕l;槣us;橱iƀ;lvỚớở䎵on»ớ;䏵ȀcsuvỪỳἋἣĀioữḱrc»Ḯɩỹ\0\0ỻíՈantĀglἂἆtr»ṝess»Ṻƀaeiἒ἖Ἒls;䀽st;扟vĀ;DȵἠD;橸parsl;槥ĀDaἯἳot;打rr;楱ƀcdiἾὁỸr;愯oô͒ĀahὉὋ;䎷耻ð䃰Āmrὓὗl耻ë䃫o;悬ƀcipὡὤὧl;䀡sôծĀeoὬὴctatioîՙnentialåչৡᾒ\0ᾞ\0ᾡᾧ\0\0ῆῌ\0ΐ\0ῦῪ \0 ⁚llingdotseñṄy;䑄male;晀ƀilrᾭᾳ῁lig;耀ffiɩᾹ\0\0᾽g;耀ffig;耀ffl;쀀𝔣lig;耀filig;쀀fjƀaltῙ῜ῡt;晭ig;耀flns;斱of;䆒ǰ΅\0ῳf;쀀𝕗ĀakֿῷĀ;vῼ´拔;櫙artint;樍Āao‌⁕Ācs‑⁒ႉ‸⁅⁈\0⁐β•‥‧‪‬\0‮耻½䂽;慓耻¼䂼;慕;慙;慛Ƴ‴\0‶;慔;慖ʴ‾⁁\0\0⁃耻¾䂾;慗;慜5;慘ƶ⁌\0⁎;慚;慝8;慞l;恄wn;挢cr;쀀𝒻ࢀEabcdefgijlnorstv₂₉₟₥₰₴⃰⃵⃺⃿℃ℒℸ̗ℾ⅒↞Ā;lٍ₇;檌ƀcmpₐₕ₝ute;䇵maĀ;dₜ᳚䎳;檆reve;䄟Āiy₪₮rc;䄝;䐳ot;䄡Ȁ;lqsؾق₽⃉ƀ;qsؾٌ⃄lanô٥Ȁ;cdl٥⃒⃥⃕c;檩otĀ;o⃜⃝檀Ā;l⃢⃣檂;檄Ā;e⃪⃭쀀⋛︀s;檔r;쀀𝔤Ā;gٳ؛mel;愷cy;䑓Ȁ;Eajٚℌℎℐ;檒;檥;檤ȀEaesℛℝ℩ℴ;扩pĀ;p℣ℤ檊rox»ℤĀ;q℮ℯ檈Ā;q℮ℛim;拧pf;쀀𝕘Āci⅃ⅆr;愊mƀ;el٫ⅎ⅐;檎;檐茀>;cdlqr׮ⅠⅪⅮⅳⅹĀciⅥⅧ;檧r;橺ot;拗Par;榕uest;橼ʀadelsↄⅪ←ٖ↛ǰ↉\0↎proø₞r;楸qĀlqؿ↖lesó₈ií٫Āen↣↭rtneqq;쀀≩︀Å↪ԀAabcefkosy⇄⇇⇱⇵⇺∘∝∯≨≽ròΠȀilmr⇐⇔⇗⇛rsðᒄf»․ilôکĀdr⇠⇤cy;䑊ƀ;cwࣴ⇫⇯ir;楈;憭ar;意irc;䄥ƀalr∁∎∓rtsĀ;u∉∊晥it»∊

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: env, file_read, file_write, network, shell.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

Remote/LAN binding instructions omit a warning about exposing presentation content and editing access.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The instructions recommend binding an editing service to all network interfaces, including an automatic Docker/WSL fallback, without requiring access controls or explaining the confidentiality and editing risks. If the service is reachable without effective authentication, other network users could access presentation content or editing functionality, potentially including the writable agent bridge.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/gateway-auth.md (reported line 33)May include surrounding context.

API_SERVER_KEY=ppt-editor-dev

text
2. Restart the Hermes Gateway (full restart, not hot reload).
3. Verify: `curl http://127.0.0.1:8642/v1/models -H "Authorization: Bearer ppt-editor-dev"`

### Default port

The troubleshooting instructions force-stop the workbench without warning about potential loss of unsaved edits.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The troubleshooting guidance recommends forcibly terminating the workbench without first saving edits or attempting graceful shutdown. Following this advice during an active editing session can lose unsaved presentation changes, creating a limited integrity and availability risk rather than an access-control exploit.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/open_workbench.py (reported line 162)May include surrounding context.

python
def _run(command: list[str], cwd: Path, timeout: int) -> subprocess.CompletedProcess[str]:
    return subprocess.run(
        command,
        cwd=str(cwd),
        text=True,

Tainted flow: 'command' from os.environ.get (line 482, credential/environment) → subprocess.run (code execution)

Medium
Category
Data Flow
Confidence
65% confidence
Finding

Data from a source is assigned to a variable that is later passed to a sink, creating a variable-mediated taint flow.

Content

Scanner excerpt · scripts/open_workbench.py (reported line 162)May include surrounding context.

python
def _run(command: list[str], cwd: Path, timeout: int) -> subprocess.CompletedProcess[str]:
    return subprocess.run(
        command,
        cwd=str(cwd),
        text=True,

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/open_workbench.py (reported line 180)May include surrounding context.

python
def _probe_command(command: str) -> tuple[bool, str]:
    """Return whether a local command is runnable, plus version/error detail."""
    try:
        result = subprocess.run(
            [command, "--version"],
            text=True,
            capture_output=True,

Opening a local presentation automatically imports unrelated Hermes gateway credentials and uses them to configure an agent bridge.

Medium
Category
Excessive Agency
Confidence
95% confidence
Finding

Opening a presentation unconditionally searches Hermes environment files and imports gateway secrets, then can automatically enable an agent transport. This expands a preview/edit request into credential-bearing agent integration without a separate authorization step and propagates those secrets to child processes.

Content

No source excerpt is available for this finding.

Launcher automatically collects agent/gateway credentials for an undeclared agent-service integration (L466–L475, L498–L508).

Medium
Category
Excessive Agency
Confidence
87% confidence
Finding

The launcher automatically selects ambient agent and gateway credentials, including the general API_SERVER_KEY fallback, and forwards them into a detached presentation server. This unnecessarily gives a long-lived runtime credentials beyond those needed for local editing; it does not by itself prove external exfiltration.

Content

No source excerpt is available for this finding.

Context Window Stuffing

Medium
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill attempts to fill the context window with filler content, displacing legitimate instructions and safety constraints. This can degrade agent performance or bypass safety boundaries.

Content

Scanner excerpt · workbench-runtime/dist-server/index.mjs (reported line 2861)May include surrounding context.

js
"cp28597": "iso88597",
      "iso88598": {
        "type": "_sbcs",
        "chars": "\x80\x81\x82\x83\x84\x85\x86\x87\x88\x89\x8A\x8B\x8C\x8D\x8E\x8F\x90\x91\x92\x93\x94\x95\x96\x97\x98\x99\x9A\x9B\x9C\x9D\x9E\x9F\xA0\uFFFD\xA2\xA3\xA4\xA5\xA6\xA7\xA8\xA9\xD7\xAB\xAC\xAD\xAE\xAF\xB0\xB1\xB2\xB3\xB4\xB5\xB6\xB7\xB8\xB9\xF7\xBB\xBC\xBD\xBE\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\u2017\u05D0\u05D1\u05D2\u05D3\u05D4\u05D5\u05D6\u05D7\u05D8\u05D9\u05DA\u05DB\u05DC\u05DD\u05DE\u05DF\u05E0\u05E1\u05E2\u05E3\u05E4\u05E5\u05E6\u05E7\u05E8\u05E9\u05EA\uFFFD\uFFFD\u200E\u200F\uFFFD"
      },
      "cp28598": "iso88598",
      "iso88599": {

Context Window Stuffing

Medium
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill attempts to fill the context window with filler content, displacing legitimate instructions and safety constraints. This can degrade agent performance or bypass safety boundaries.

Content

Scanner excerpt · workbench-runtime/dist-server/index.mjs (reported line 3042)May include surrounding context.

js
"csibm1133": "cp1133",
      "cp1161": {
        "type": "_sbcs",
        "chars": "\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\uFFFD\u0E48\u0E01\u0E02\u0E03\u0E04\u0E05\u0E06\u0E07\u0E08\u0E09\u0E0A\u0E0B\u0E0C\u0E0D\u0E0E\u0E0F\u0E10\u0E11\u0E12\u0E13\u0E14\u0E15\u0E16\u0E17\u0E18\u0E19\u0E1A\u0E1B\u0E1C\u0E1D\u0E1E\u0E1F\u0E20\u0E21\u0E22\u0E23\u0E24\u0E25\u0E26\u0E27\u0E28\u0E29\u0E2A\u0E2B\u0E2C\u0E2D\u0E2E\u0E2F\u0E30\u0E31\u0E32\u0E33\u0E34\u0E35\u0E36\u0E37\u0E38\u0E39\u0E3A\u0E49\u0E4A\u0E4B\u20AC\u0E3F\u0E40\u0E41\u0E42\u0E43\u0E44\u0E45\u0E46\u0E47\u0E48\u0E49\u0E4A\u0E4B\u0E4C\u0E4D\u0E4E\u0E4F\u0E50\u0E51\u0E52\u0E53\u0E54\u0E55\u0E56\u0E57\u0E58\u0E59\u0E5A\u0E5B\xA2\xAC\xA6\xA0"
      },
      "ibm1161": "cp1161",
      "csibm1161": "cp1161",