Back to skill

Security audit

acpx-faq

Security checks for vulnerabilities and agentic risk

Overview

This coding-agent guide needs review because it advises bypassing approval rules and prompt screening, and its setup command can overwrite existing settings.

Review before installing. Remove the approval-override and screening-bypass instructions, and preserve existing settings when configuring authentication. Authorize a specific unattended task, use an enforced sandbox, and verify the provider, account, project access, and inherited credentials. Close failed sessions and consider retained logs. The partial scanner report was checked against both bundled files; external executables and your runtime configuration were not audited.

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
Findings (1)

T01 · Skill Instruction Hijacking

Warning
Location
SKILL.md
Finding

Explicit instructions to evade harness screening and inherited approval requirements

Content
View full analysis

Impact

An agent following these instructions is directed to conceal blocked content from command-level screening and suppress required approval questions, potentially allowing edits or delegated operations that the existing controls would stop. The files demonstrate the bypass directive, not successful execution against a particular harness.

Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Rogue AgentSelf-Modification, Session Persistence
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (16)

Self-Modification

High
Category
Rogue Agent
Confidence
90% confidence
Finding

Skill modifies its own code, configuration, or behavior at runtime. Self-modification enables an agent to escalate privileges, disable safety constraints, or install persistent backdoors.

Content

Scanner excerpt · CHANGELOG.md (reported line 336)May include surrounding context.

md
- agy: provenance of `agy_acp_server.par` - shipped by the Antigravity install, sha512-verified
  against Google's manifest, never downloaded by acpx or by this skill. Build
  `20260818_01_RC01` is pinned and self-updates nothing. Clears the Socket alert on unverified
  executable provenance.
- agy: Linux / NixOS section - the `--uid=` wrapper, the mandatory `SSL_CERT_FILE` export, and
  installing `agy` by hand rather than piping the installer (its `agy install` step writes to

Headless-operation guidance instructs executors to override inherited user confirmation requirements.

High
Category
Prompt Injection
Confidence
97% confidence
Finding

The skill identifies an inherited user rule such as 'ask before editing' and prescribes a blanket instruction to act without asking. Headless execution does not authorize overriding user confirmation requirements, particularly when the recipes also grant broad approvals.

Content

No source excerpt is available for this finding.

External Model or Provider Selection

High
Category
Excessive Agency
Confidence
90% confidence
Finding

Skill selects an external model or provider that may use a different account or billing plan than the operator expects. Undisclosed model switches can cause unexpected cost or quota consumption.

Content

Scanner excerpt · SKILL.md (reported line 150)May include surrounding context.

bash
D=/abs/real/dir                              # must exist: roots resolve through realpath
acpx --agent ~/.local/lib/antigravity-acp/agy_acp_server.par \
     --cwd "$D" --model gemini-3.8-flash --approve-all --timeout 1800 \
     exec "Carry out $D/brief.md. Write your report to $D/report.md."

External Model or Provider Selection

High
Category
Excessive Agency
Confidence
90% confidence
Finding

Skill selects an external model or provider that may use a different account or billing plan than the operator expects. Undisclosed model switches can cause unexpected cost or quota consumption.

Content

Scanner excerpt · SKILL.md (reported line 257)May include surrounding context.

bash
D=/abs/dir
acpx --cwd "$D" --model claude-opus-5 claude sessions ensure --name work
acpx --cwd "$D" claude set effort high -s work                 # optional; after any model change
{ acpx --cwd "$D" --model claude-opus-5 --approve-all --suppress-reads --timeout 2400 \
     claude -s work -f /abs/brief.md && \

External Model or Provider Selection

High
Category
Excessive Agency
Confidence
90% confidence
Finding

Skill selects an external model or provider that may use a different account or billing plan than the operator expects. Undisclosed model switches can cause unexpected cost or quota consumption.

Content

Scanner excerpt · SKILL.md (reported line 259)May include surrounding context.

md
D=/abs/dir
acpx --cwd "$D" --model claude-opus-5 claude sessions ensure --name work
acpx --cwd "$D" claude set effort high -s work                 # optional; after any model change
{ acpx --cwd "$D" --model claude-opus-5 --approve-all --suppress-reads --timeout 2400 \
     claude -s work -f /abs/brief.md && \
  acpx --cwd "$D" claude sessions close work; } >> run.log 2>&1 &
# failed turn (exit 3/5) = close skipped: re-prompt -s work to recover with context,

The skill instructs agents to bypass driving-harness prompt controls by submitting blocked content through files.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

The skill explicitly recommends file-based delivery when a driving harness blocks prompt text through a hook or classifier. This attempts to route rejected instructions around a security boundary into an executor with tool access; legitimate file-based prompting does not justify bypassing a denial.

Content

No source excerpt is available for this finding.

File-based prompt delivery is recommended specifically to evade the driving harness's safety checks.

High
Category
Prompt Injection
Confidence
99% confidence
Finding

This is the same explicit control-evasion instruction assessed under SDI-2: content blocked by the driving harness is to be submitted through a file instead. It crosses the harness-to-executor trust boundary by changing transport rather than resolving the reason for rejection.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx fallback'.

Medium
Category
MCP Rug Pull
Confidence
70% confidence
Finding

npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.

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 59)May include surrounding context.

md
on it.
4. **Permission flags are not a sandbox.**
   - **claude**: the adapter starts in `permissions.defaultMode` from the Claude
     settings (`acceptEdits` / `auto` approve edits without asking). Only what that
     mode escalates reaches acpx as a permission request - `--deny-all` refuses those
     (exit 5), but an auto-approved edit never asks.
   - **codex**: writes inside its workspace-write sandbox (rooted at `--cwd`, plus temp

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
85% 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 61)May include surrounding context.

md
- **claude**: the adapter starts in `permissions.defaultMode` from the Claude
     settings (`acceptEdits` / `auto` approve edits without asking). Only what that
     mode escalates reaches acpx as a permission request - `--deny-all` refuses those
     (exit 5), but an auto-approved edit never asks.
   - **codex**: writes inside its workspace-write sandbox (rooted at `--cwd`, plus temp
     dirs) raise no request, so `--deny-all` still writes; outside fails with
     `Operation not permitted` - including a linked worktree's gitdir, so a sandboxed

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 71)May include surrounding context.

md
at spawn, or `codex set-mode agent-full-access -s <n>` on a live session, does.
   - `--allowed-tools` / `--max-turns` reach only claude (and qoder); codex and agy
     ignore them. `--allowed-tools` pre-approves and never restricts: `--help`'s
     `"" for no tools` means no approvals, not no tools.

   The real filesystem control is which `--cwd` you hand it.
5. **`[done] end_turn` and exit 0 are not proof of success.** A content-filter kill,

MCP server referenced without pinned version: 'npx cache'.

Medium
Category
MCP Rug Pull
Confidence
70% confidence
Finding

npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.

Content

No source excerpt is available for this finding.

Agent-dispatch instructions omit disclosure that task content may be sent to external model providers.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The workflow instructs an authenticated Gemini coding agent to read and execute a local brief, but does not warn that the brief and accessed project content may be transmitted to the external model provider. Filesystem, credential-inheritance, and permission warnings elsewhere do not explain this external data-processing exposure.

Content

No source excerpt is available for this finding.

The authentication setup command overwrites any existing Antigravity settings without warning.

Medium
Category
Output Handling
Confidence
98% confidence
Finding

The shell redirection truncates an existing settings.json and replaces it with only the authentication setting, without checking or backing it up. Following this setup recipe can erase unrelated configuration, potentially including security-relevant settings, and disrupt an existing installation.

Content

No source excerpt is available for this finding.

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 · SKILL.md (reported line 172)May include surrounding context.

  • Auth: needs auth.type in <GEMINI_HOME>/antigravity-acp/settings.json (one-time):

    bash
    mkdir -p ~/.gemini/antigravity-acp
    echo '{"auth":{"type":"oauth-personal"}}' > ~/.gemini/antigravity-acp/settings.json
    

MCP server referenced without pinned version: 'npx cache'.

Medium
Category
MCP Rug Pull
Confidence
70% confidence
Finding

npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.

Content

No source excerpt is available for this finding.