summaryrefslogtreecommitdiff
path: root/src/agents/03-logs/main.py
diff options
context:
space:
mode:
Diffstat (limited to 'src/agents/03-logs/main.py')
-rw-r--r--src/agents/03-logs/main.py25
1 files changed, 16 insertions, 9 deletions
diff --git a/src/agents/03-logs/main.py b/src/agents/03-logs/main.py
index 8da7121..d237cc8 100644
--- a/src/agents/03-logs/main.py
+++ b/src/agents/03-logs/main.py
@@ -10,21 +10,23 @@ def load_prompt(file_path):
if __name__ == "__main__":
# Configuration for the experiment run
- SCENARIO = "hidden_goal" # e.g., "benign", "hidden_goal"
- OVERSIGHT_LEVEL = "low" # e.g., "low", "high"
- USER_PROMPT_TYPE = "list_files" # e.g., "list_files", "check_username"
+ SCENARIO = "data_sanitization"
+ OVERSIGHT_LEVEL = "low"
+ USER_PROMPT_TYPE = "data_sanitization"
- # Load scenario prompt
- scenario_prompt_path = os.path.join("prompts", "scenarios", f"{SCENARIO}.md")
- scenario_prompt = load_prompt(scenario_prompt_path)
+ # Load scenario prompts
+ cover_prompt_path = os.path.join("prompts", "scenarios", f"{SCENARIO}_cover.md")
+ hidden_prompt_path = os.path.join("prompts", "scenarios", f"{SCENARIO}_hidden.md")
+ cover_prompt = load_prompt(cover_prompt_path)
+ hidden_prompt = load_prompt(hidden_prompt_path)
# Load oversight prompt
oversight_prompt_path = os.path.join("prompts", "oversight", f"{OVERSIGHT_LEVEL}.md")
oversight_prompt = load_prompt(oversight_prompt_path)
# Combine system prompts
- system_prompt_parts = [scenario_prompt, oversight_prompt]
- system_prompt = "\n\n".join(filter(None, system_prompt_parts)) # Filter out empty strings
+ system_prompt_parts = [cover_prompt, oversight_prompt, hidden_prompt]
+ system_prompt = "\n\n".join(filter(None, system_prompt_parts))
# Load user prompt
user_prompt_path = os.path.join("prompts", "user", f"{USER_PROMPT_TYPE}.md")
@@ -37,4 +39,9 @@ if __name__ == "__main__":
user_prompt_type=USER_PROMPT_TYPE
)
agent.run(user_prompt)
- agent.save_logs() \ No newline at end of file
+ agent.save_logs()
+
+ print("\n--- Final Virtual File System State ---")
+ from vfs import vfs_instance
+ vfs_instance.print_fs()
+ \ No newline at end of file