fix: security and crash bugs in shell, crypto, browser, content skills

[lbartoszcze]

Summary

Fixes 2 security vulnerabilities and 2 crash bugs in the agent runtime, with 22 tests.

1. shell.py — _spawn() bypasses all command safety checks (CRITICAL)

The bug: _bash() checks commands against a dangerous-command blocklist before execution. _spawn() does NOT — it passes commands directly to subprocess.Popen(command, shell=True). Any agent can bypass all safety by using shell:spawn instead of shell:bash to run destructive commands as background processes.

The fix: Extracted the blocklist check into _check_dangerous_command() and apply it to both _bash() and _spawn().

2. crypto.py — Private keys leaked in SkillResult.data (CRITICAL)

The bug: _create_wallet() returns the private key in SkillResult.data["_private_key"]. This data flows into:

• recent_actions → included in LLM prompts → sent to third-party API providers
• activity.json → written to disk in plaintext

Any agent holding real funds has its keys exposed to LLM providers and log files.

The fix: Removed _private_key from result data. The key remains stored securely in self._wallets for actual blockchain operations.

3. browser.py — AttributeError on cleanup (HIGH)

The bug: __init__() initializes self.browser, self.context, self.page but never self._playwright. When _close_browser() runs before _ensure_browser() (e.g., during shutdown), it crashes with AttributeError: 'BrowserSkill' object has no attribute '_playwright'.

The fix: Initialize self._playwright = None in __init__(). Also guard _close_browser() to handle _playwright and browser independently.

4. content.py — AttributeError when no API key available (MEDIUM)

The bug: _init_llm() sets self.model in the Anthropic and OpenAI branches, but the else branch (no API key) only sets self.llm and self.llm_type, leaving self.model undefined. Downstream code that accesses self.model crashes with AttributeError.

The fix: Set self.model = None in the else branch.

Files changed

File	Change
singularity/skills/shell.py	Extract _check_dangerous_command(), apply to _spawn()
singularity/skills/crypto.py	Remove _private_key from SkillResult.data
singularity/skills/browser.py	Initialize self._playwright = None, guard cleanup
singularity/skills/content.py	Set self.model = None in else branch
pyproject.toml	Add pytest-timeout to dev deps, pytest config
.github/workflows/ci.yml	CI workflow (lint + test + build)
tests/test_security_and_crash_fixes.py	22 tests covering all 4 fixes

Test plan

• 22 tests pass locally (pytest tests/ -v --timeout=30)
• Lint passes on modified files
• Package builds cleanly

🤖 Generated with Claude Code