Research Strategy
PassAudited by VirusTotal on May 12, 2026.
Overview
Type: OpenClaw Skill Name: research-strategy Version: 1.0.0 The skill bundle is classified as suspicious due to explicit prompt injection instructions in SKILL.md that command the AI agent to operate autonomously without user confirmation ('无需等待确认', '直接执行', '循环', '自动继续下一个策略'). This bypasses critical human oversight. Additionally, the `research_workflow.py` script directly modifies Python strategy files via string replacement in its `reverse_logic` function, which is a high-risk operation that could lead to unintended code changes or breakage, even if its stated purpose is to adjust trading logic.
Findings (0)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
It may keep searching, coding, backtesting, and changing files after the user expected a single bounded task.
The skill explicitly tells the agent to spawn a sub-agent that runs autonomously, does not wait for confirmation, and loops through strategy research without a clear stop condition.
sessions_spawn(... label="Research Agent") ... "不需要等确认，直接执行" ... "自动继续下一个策略" ... "循环直到所有策略研究完" ... "自主执行，无需等待确认。"
Require explicit user opt-in before spawning sub-agents or background workers, add a finite iteration/time limit, and provide a clear stop command.
A flawed generated strategy could be promoted into the user's trading project and later be used as if it were validated.
The script can automatically move a generated test strategy into the formal strategy directory based only on backtest metrics, with no manual approval or review gate.
if decision == 'MOVE_TO_FORMAL':
if move_to_formal(result['strategy']):
log(f"✅ 已移到正式文件夹")
...
os.rename(src, dst)Require manual review before promoting strategies, create backups, validate paths, and keep generated strategies isolated until explicitly approved.
Unreviewed local code may run in the user's trading workspace and continue consuming resources or changing project state in the background.
The workflow directs background execution of local Python scripts, including backtests over newly created strategy code, while the same document says confirmation is not needed.
python3 BackTest_Research-strategy.py & ... python3 research_workflow.py &
Run generated strategy code only in a sandbox, avoid background execution by default, and ask for confirmation before each execution step.
Bad report data or a faulty generated strategy can propagate into code files and persistent project memory without containment.
A single backtest report can trigger multiple persistent changes: promoting a strategy, rewriting its logic, and recording the result into memory.
if decision == 'MOVE_TO_FORMAL':
... move_to_formal(result['strategy'])
elif decision == 'REVERSE_LOGIC':
... reverse_logic(result['strategy'])
# 记录
record_to_memory(result['strategy'], result, decision, reason)Validate reports, stage changes for review, add rollback support, and separate experimental outputs from formal project state.
Incorrect or crafted report content could become persistent context that future agent runs may trust.
The script appends report-derived strategy information directly into MEMORY.md, a persistent context file, without sanitizing or requiring review.
entry = f"""
## {strategy_name}（{timestamp}）
...
"""
...
with open(MEMORY_FILE, 'a') as f:
f.write(entry)Treat MEMORY.md updates as reviewable data, sanitize report-derived text, and avoid storing executable instructions or untrusted content in persistent memory.