Back to skill

Security audit

riffkit

Security checks for vulnerabilities and agentic risk

Overview

ClawScan rates this video creation and editing skill suspicious because it reports automatic replacement of its own instructions with remote content, along with other behavior beyond the stated purpose; analysis was partial.

The verdict is suspicious, with medium overall confidence, and the scanner recommends not installing. Consider the reported remote instruction updates and retained session token before deciding; inspection was incomplete because references/api.md was only partially evaluated.

Vulnerability Patterns
  • 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
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Memory PoisoningPersistent Context Injection, Context Window Stuffing, Memory Manipulation
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
Findings (24)

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · HEARTBEAT.md (reported line 35)May include surrounding context.

bash
# Remote version (via the /SKILL.json endpoint; timestamp the URL to dodge caches)
REMOTE_VERSION=$(curl -s "${BASE_URL}/SKILL.json?t=$(date +%s)" \
  | node -e 'let s="";process.stdin.on("data",d=>s+=d).on("end",()=>process.stdout.write(JSON.parse(s).version))')

# Local version (extracted from SKILL.md's frontmatter)

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · HEARTBEAT.md (reported line 45)May include surrounding context.

md
> **No `node`?** Use `python3`:
> ```bash
> REMOTE_VERSION=$(curl -s "${BASE_URL}/SKILL.json?t=$(date +%s)" | python3 -c 'import json,sys;print(json.load(sys.stdin)["version"])')
> ```

**Comparison rule:** a plain **string equality** check (Riffkit version numbers are always minted by the server; the local copy is never newer than remote).

The heartbeat automatically replaces the agent's installed skill instructions with remote content, a capability outside the stated video-generation purpose.

High
Category
Excessive Agency
Confidence
97% confidence
Finding

The update procedure overwrites installed agent instructions and references with mutable remote content without review, explicit update approval, or release integrity verification. Compromise of the distribution endpoint could therefore persist attacker-controlled instructions that inherit the agent's available capabilities; HTTPS alone does not protect against a compromised publisher.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The supplied chunk implements a skill-content updater rather than video generation or editing. It overwrites SKILL.md and supporting documents from BASE_URL, enabling remote changes to the skill's operational instructions—an undeclared capability distinct from processing video inputs or invoking a video-generation backend. The chunk does not itself execute downloaded content or demonstrate data exfiltration.

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 17)May include surrounding context.

md
3. **No Node 20+, or the install fails?** Each command is one HTTP call: references/api.md lists them, starting with **Auth**.

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 41)May include surrounding context.

md
names its `riffkit` command; without the CLI, make the HTTP call it stands for (references/api.md, "Commands and routes"). A command that spends credits runs on

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 55)May include surrounding context.

md
's length in seconds (without the CLI, send it as `video` on `POST /api/riffs`: references/api.md).

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 103)May include surrounding context.

md
Before the burn, `riffkit preview_subtitles` renders a free preview (references/api.md, "Subtitle editing").

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 125)May include surrounding context.

md
- On **HTTP 402**, follow "Billing & balance" (references/api.md): relay `topup_url` verbatim, **no retry, no silent failure**

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 157)May include surrounding context.

md
- [references/api.md](references/api.md) (https://riffkit.ai/skill/api.md): API reference

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
80% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · references/api.md (reported line 45)May include surrounding context.

md
| `riffkit save_subtitles` | `PUT /api/assets/{asset_id}/subtitles` | Save subtitle edits |
| `riffkit preview_subtitles` | `POST /api/assets/{asset_id}/subtitles/preview` | A free preview of the edited subtitles |
| `riffkit burn_subtitles` | `POST /api/assets/{asset_id}/subtitles/burn` | Render the edits into a new version of the video |
| `riffkit reset_subtitles` | `DELETE /api/assets/{asset_id}/subtitles/edits` | Undo the edits |
| `riffkit reconcile_subtitles` | `POST /api/assets/{asset_id}/subtitles/reconcile` | Rebuild a video's subtitle lines |
| `riffkit analyze_template` | `POST /api/formulas/analyze` | Subscribers only: analyze a new source into a template, without making a video |
| `riffkit refresh_template` | `POST /api/formulas/{formula_id}/refresh-analysis` | Re-analyze one of your own templates |

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
80% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · references/api.md (reported line 724)May include surrounding context.

md
**Body:** `{entities: [...]}` — the complete replacement list, checked against the burn contract. A 400 means an entity has the wrong shape (a `kind` other than `"subtitle"` / `"graphic"`, a missing `id` / `semantic` / `time_range`, a wrong type) or a graphic layer's image key changed, and says what to fix. Style values the burn can't use are removed instead of rejected: an unrecognized `color` / `highlight_color`, a `highlight_words` that isn't a list of strings, or entries not found in `params.text`; compare the returned `entities` with what you sent. `params.font` isn't supported (there is no per-line font) and is removed on save. Other values (e.g. `approximate_size`) aren't checked here, so stick to the listed options. Saving does NOT change the video — only `burn` does.

#### `DELETE /api/assets/{asset_id}/subtitles/edits`

Reset to the machine baseline. Idempotent; returns `{reset, source}`.

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
80% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · references/api.md (reported line 51)May include surrounding context.

md
| `riffkit refresh_template` | `POST /api/formulas/{formula_id}/refresh-analysis` | Re-analyze one of your own templates |
| `riffkit update_template` | `PATCH /api/formulas/{formula_id}` | Rename one of your own templates or change its hint |
| `riffkit set_voice_sample` | `POST /api/characters/{character_id}/voice-sample` | Set a character's voice sample |
| `riffkit clear_voice_sample` | `DELETE /api/characters/{character_id}/voice-sample` | Remove a character's voice sample |
| `riffkit get_credits` | `GET /api/usage/credits` | The balance, today's spend and the daily limit |
| `riffkit get_daily_budget` | `GET /api/usage/daily-budget` | Today's spend against the daily limit |
| `riffkit get_usage_summary` | `GET /api/usage/summary` | Spend summary |

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
80% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · references/api.md (reported line 486)May include surrounding context.

md
**Content-Type:** `multipart/form-data`, field `audio` (**mp3/wav only** — Seedance accepts exactly these; ≤5MB, duration 4-15s — 5-10s is best; no background music/noise). **Response:** updated `CharacterOut`. Replacing = upload again (pointer swaps).

#### `DELETE /api/characters/{character_id}/voice-sample`

Clears the sample (generation falls back to the default voice). **Response:** updated `CharacterOut`.

Anti-Refusal Statement

High
Category
Anti-Refusal
Confidence
85% confidence
Finding

Skill instructs the agent to omit warnings, disclaimers, or ethical commentary. Stripping safety caveats hides risk from the user and is a common jailbreak preamble.

Content

Scanner excerpt · references/api.md (reported line 683)May include surrounding context.

md
### Subtitle editing (post-production, free)

Fix a finished video's subtitles without regenerating it: retime a line, move captions out of a face, change text/color/size, delete a line, then re-burn. **Zero-charge** — burn/reconcile are pure post-production (no video generation), so no credits are ever spent here; don't warn the user about cost. All endpoints take the **asset id** of the finished video (`asset_role=final_reel`). A video made in sections whose `staged.pending` is true can be read but not changed: `PUT`, `DELETE .../edits`, `burn` and `reconcile` answer 409 `film_pending` (its captions belong to the whole video) until all of it is made.

**What you can edit, on every video:** the caption layer Riffkit burns on top of the picture, and nothing else; `GET` returns that layer. Text that is part of the picture itself can't be edited, moved or removed here. On a swap (`type: "swap"`) that is the common case: the source video's on-screen text (its captions, labels, stickers)

Memory Manipulation

High
Category
Memory Poisoning
Confidence
85% confidence
Finding

Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.

Content

Scanner excerpt · references/details.md (reported line 12)May include surrounding context.

md
emotion formula: hook, rhythm, beats | The story, scenes, script, language | The user wants *their own* video that works like the winner |
| **Swap** | `swap` | The source's camera, cuts, framing, action, timing, sound and frame shape | What the user names: the person (your character, if you pick one), a product, and whatever `content_anchor` names: the setting, an outfit, a line's wording | The user wants *this* video with their character in it ("the same video, but me", "keep every shot") |

Swap rules (backend-enforced):
- **A swap must change at least one thing**: a character (`character_ids`), a product that has images (`product_id`; a product without images changes nothing), a reference image (`reference_ids`; an outfit needs a person in the source), or a non-empty `content_anchor`. None of these → 400 whose `detail` is a plain localized sentence (en: "A swap needs at least one change: …"; the body carries no error code, so relay `detail` rather than matching on it). Checked for

The download procedure contradicts the documented promise to keep old files when an update fails.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

Shell output redirection truncates each installed file before curl succeeds, contradicting the promise to preserve old files on failure. Network errors, HTTP error bodies, or interrupted downloads can leave the agent with missing, corrupted, or mixed-version instructions and safety constraints.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 49)May include surrounding context.

md
- **Adapt** keeps the formula: the original's hook and rhythm, telling the user's own story with new footage.
- **Swap** keeps the shots: the original's camera, cuts, action, timing, sound, language and frame shape, with only who or what is in them changed.

Take the mode without asking only when the user says Adapt or Swap, or states the choice itself: "keep every shot", "the same video with me in it" = Swap; "a new story", "new footage, my own script" = Adapt. A request that only names a product, a character, a link or a template fits both modes: ask before any price: *Keep the original shots and change who or what is in them (Swap), or keep what made it work and tell your own story (Adapt)?* Send `mode` on every `riffkit quote_remake` and `riffkit remake_video` call: left out, the call runs as adapt.

**1. Source: the one required input, exactly one.**
- A template: `riffkit list_templates` (status analyzed; prefer public and most-used ones), and `riffkit get_template` to see what o

External Transmission

Medium
Category
Data Exfiltration
Confidence
70% 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/api.md (reported line 78)May include surrounding context.

md
Every path below already includes the full prefix — just append it to `${BASE_URL}` (e.g. `GET /api/auth/me` → `https://riffkit.ai/api/auth/me`).

⚠️ **Request bodies must be UTF-8.** Python `requests.post(url, json=...)`, Node `fetch`/`axios`, Go `json.Marshal` are UTF-8 by default — pure-ASCII needs nothing. **Only** on Chinese Windows `cmd` run `chcp 65001` first (PowerShell also needs `[Console]::OutputEncoding = [System.Text.Encoding]::UTF8`), or non-ASCII characters get sent as GBK and rejected with `BAD_REQUEST`. Never assemble a byte string with `data=` in any language.

### Auth

External Transmission

Medium
Category
Data Exfiltration
Confidence
70% 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/errors.md (reported line 38)May include surrounding context.

md
send the same command: it fails the same way, and `chcp` alone doesn't fix it. Write the text to a UTF-8 file and pass it by reference: `-F "content_anchor=<'/tmp/anchor.txt'"` (see `POST /api/riffs`). Already doing that? Rewrite the file with an explicit UTF-8 encoding. Never switch a multipart endpoint to `json=`: its fields would be ignored |
| `400` — "There was an error parsing the body" | a JSON body (e.g. `POST /api/products`) was not UTF-8 | Send JSON with your library's UTF-8 encoder (`requests.post(url, json=…)`, `fetch`/`axios`, `json.Marshal`); on Chinese Windows `cmd` run `chcp 65001` first; never build the body by hand with `data=` |
| `422` — request validation | a JSON body without a required field (e.g. `creation/batch` without `content_anchor`, or with it empty or only spaces); a wrong type or an unsupported value (e.g. `product_visibility` / `duration_mode` outside the listed values, `duration_seconds` outside 4-45, `duration_mode=fixed` without `duration_seconds`);

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · references/api.md (reported line 103)May include surrounding context.

  • stop after expires_in (10 min) and tell the user the link expired
  1. Keep the token for the whole session. Each shell command runs in a new process, so a token held in a shell variable is gone after that one command, and the next call would need a new sign-in. Right after approved, save it to a private file and read it back in every later command:
    bash
    mkdir -p ~/.riffkit && (umask 077 && printf '%s' "$TOKEN" > ~/.riffkit/session)
    curl -sS -b "vee_session=$(cat ~/.riffkit/session)" "https://riffkit.ai/api/auth/me"
    
    Reuse it until a request returns 401, then delete the file and run the device flow again. Never print it. In zsh, don't name a polling variable status: it is read-only there and breaks the loop.

The skill exposes subscription cancellation and renewal controls beyond its stated video-creation purpose.

Medium
Category
Excessive Agency
Confidence
96% confidence
Finding

The manifest describes generating and modifying short videos, but these routes let the agent cancel a paid subscription or restore automatic renewal. Unlike checking credits or quoting render costs, changing the account's subscription lifecycle is not necessary for video creation and can affect future charges or end an unpaid plan immediately; the documented confirmation requirement does not remove this scope mismatch.

Content

No source excerpt is available for this finding.

The standalone trigger 'riff' can activate the skill outside video-generation requests.

Low
Category
Trigger Abuse
Confidence
93% confidence
Finding

The trigger list explicitly includes 'riff' without requiring video-related context. This word also occurs in music and general brainstorming requests, and no exclusion condition prevents those uses from activating the video skill.

Content

No source excerpt is available for this finding.

Scope Creep

Low
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill's behavior or capabilities extend beyond its stated purpose. Scope creep allows an agent to perform actions unrelated to its documented functionality, increasing the attack surface.

Content

Scanner excerpt · SKILL.md (reported line 144)May include surrounding context.

md
6. **Confirm the file's purpose before uploading**, to avoid uploading sensitive documents by mistake.
7. **Never publish on the user's behalf** to any external platform — the output is local material; publishing rights are the user's.
8. **Never fabricate data**: this skill provides no performance metrics (no TikTok data endpoints); if asked, say it's unavailable rather than inventing it.
9. **Don't expand scope**: only call the endpoints listed here; don't probe other paths or call staff/admin endpoints.
10. **Don't read/write unrelated local files**: only in a context the user explicitly requested (e.g. "upload this product image /path/x.jpg").

Proactively flag anomalies (an undocumented error code / an internal field that shouldn't be exposed / the same task failing after 2 retries / balance dropping >10% in a minute for no reason).