Problem Statement
When running autonomous code generation with AutoGen, UserProxyAgent and LocalCommandLineCodeExecutor execute generated code blocks directly on the host machine or container.
Current approaches to mitigate dangerous executions either:
- Wrap execution in heavy virtualization/Docker containers (high resource overhead, cold start latency).
- Use remote LLM evaluation proxies (adding 1,200ms to 2,500ms latency per code step).
Proposed Solution: Zero-Latency In-Memory AST Gating
We have built btp-guard (Bartholomew Trust Protocol), an open-source, deterministic compiler-level AST invariant gate that inspects proposed Python, Bash, and SQL code blocks directly in memory in under 50 microseconds before dispatch.
Drop-In AutoGen Integration Example
It plugs directly into AutoGen's custom code execution pipeline:
from autogen.coding import LocalCommandLineCodeExecutor
from btp_guard import guard, SecurityVetoException
class SecureCodeExecutor(LocalCommandLineCodeExecutor):
def execute_code_blocks(self, code_blocks):
for block in code_blocks:
# Sub-50 microsecond in-memory AST and credential leak verification
is_safe, reason = guard.evaluate(block.language, block.code)
if not is_safe:
return 1, f"[BTP Security Veto] Execution aborted: {reason}", None
return super().execute_code_blocks(code_blocks)Key Technical Properties:
- Execution Overhead: <50 microseconds (over 30,000x faster than cloud moderation proxies).
- Attack Coverage: Catastrophic filesystem wipes (
rm -rf), DDL SQL purges (DROP TABLE), fork bombs, dynamic Python introspection (__subclasses__), and high-entropy secret leaks.
- Availability: Published on PyPI (
pip install btp-guard) and npm (npm install btp-guard).
- Attestation: RFC 8785 canonical JSON Ed25519 cryptographic receipts for audit logging.
Would the AutoGen maintainers be interested in a cookbook example or extension recipe demonstrating pre-flight in-memory invariant gating for secure agent code execution?
Problem Statement
When running autonomous code generation with AutoGen,
UserProxyAgentandLocalCommandLineCodeExecutorexecute generated code blocks directly on the host machine or container.Current approaches to mitigate dangerous executions either:
Proposed Solution: Zero-Latency In-Memory AST Gating
We have built
btp-guard(Bartholomew Trust Protocol), an open-source, deterministic compiler-level AST invariant gate that inspects proposed Python, Bash, and SQL code blocks directly in memory in under 50 microseconds before dispatch.Drop-In AutoGen Integration Example
It plugs directly into AutoGen's custom code execution pipeline:
Key Technical Properties:
rm -rf), DDL SQL purges (DROP TABLE), fork bombs, dynamic Python introspection (__subclasses__), and high-entropy secret leaks.pip install btp-guard) and npm (npm install btp-guard).Would the AutoGen maintainers be interested in a cookbook example or extension recipe demonstrating pre-flight in-memory invariant gating for secure agent code execution?