Adversarial AI Cyber Threat Intelligence Swarm with AutoGen and Baseten
State-sponsored cyber actors are leveraging offensive AI tools like DeepSeek and Cursor to target corporate networks. You will construct a multi-agent conversational threat intelligence swarm using AutoGen hosted on Baseten high-performance inference endpoints. The agent swarm must synthesize multi-source threat telemetry and produce proactive firewall rules with 95%+ precision.
What you are building
The core problem, expected build, and operating context for this challenge.
Build an AutoGen multi-agent conversational swarm running on Baseten infrastructure to detect and neutralize AI-powered cyber attack patterns in real time.
How work is evaluated
Evaluate the defensive agent group's ability to identify deepseek/cursor malware logs and generate valid YARA signatures.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
yara_syntax_check
Verifies that generated YARA rule contains required rule keywords.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
threat_precision
Precision of threat identification against MITRE ATT&CK taxonomy. • target: 0.95 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Design AutoGen `AssistantAgent` and `UserProxyAgent` instances for threat hunting and policy synthesis.
Deploy custom threat classification models to Baseten Serverless for sub-100ms inference.
Implement multi-turn conversation loops between Red Team attacker simulators and Blue Team defense agents.
Produce automated, valid Snort and YARA defensive rules from raw network event logs.
Reference links and supporting material
Synthetic telemetry dataset of 300 network traffic logs and memory dumps representing AI-generated attack payloads.
Challenge setup
Evaluate the defensive agent group's ability to identify deepseek/cursor malware logs and generate valid YARA signatures.
Challenge input
JSON containing network_log, attack_vector, suspected_ai_tool, compromised_asset
AutoGen
Enables multi-agent conversational problem solving between defense tools.
Baseten
Provides fast, dedicated GPU endpoints for custom security analysis models.
Evaluated output
JSON containing identified_threat, yara_rule, snort_rule, threat_severity
- Verifies that generated YARA rule contains required rule keywords.
- Precision of threat identification against MITRE ATT&CK taxonomy. • target: 0.95 • range: 0-1
- Benchmark: AICyberDefBench-2026
- Threat Precision target: 0.95
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
- Baseten · Required
- AutoGen · Optional
- DeepSeek · Optional
Evaluation contract
- yara_syntax_check · Weight 1
- threat_precision · Weight 1
Metrics
- threat_precision · Target 0.95
Benchmarks
- AICyberDefBench-2026
- security
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON containing network_log, attack_vector, suspected_ai_tool, compromised_asset
Output: JSON containing identified_threat, yara_rule, snort_rule, threat_severity
Execution harness: Python
Evaluation tools
- AutoGen · Multi-Agent Framework
- Baseten · Inference Infrastructure
Recipe freeze marker
No recipe freeze time is recorded.