AI & Agentic Workflows in Cybersecurity
The integration of Artificial Intelligence and Large Language Models (LLMs) into cybersecurity operations accelerates log analysis, scan triage, and remediation reporting. However, deploying AI in security requires strict adherence to data confidentiality, offline execution, and human-in-the-loop control.
Local Inference Architecture with Ollama
Section titled “Local Inference Architecture with Ollama”flowchart LR ScanLogs["Raw Nmap / Telemetry Logs"] --> Parser["Sanitization & Preprocessing"] Parser --> Ollama["Local Inference Server (Ollama / 127.0.0.1:11434)"] Ollama --> Model["Local Quantized Model (Qwen2.5-Coder / Llama-3)"] Model --> JSON["Validated Structured JSON Output"] JSON --> Human["Human Reviewer & Report Integration"]Deploying Local Inference Engine
Section titled “Deploying Local Inference Engine”# 1. Install Ollama local runtimecurl -fsSL https://ollama.com/install.sh | sh
# 2. Pull a high-capability cybersecurity & code reasoning modelollama pull qwen2.5-coder:7b
# 3. Confirm model availability and resource allocationollama listAutomated Vulnerability Triage with Python
Section titled “Automated Vulnerability Triage with Python”Utilizing Python to connect directly to the local Ollama REST API (/api/generate) with guaranteed JSON output formatting:
#!/usr/bin/env python3import jsonimport requests
OLLAMA_ENDPOINT = "http://127.0.0.1:11434/api/generate"MODEL_NAME = "qwen2.5-coder:7b"
def triage_vulnerability(service_banner: str) -> dict: prompt = ( f"You are a defensive cybersecurity analyst. Given this service banner:\n" f"'{service_banner}'\n\n" f"Return a strict JSON object with these keys:\n" f"- 'cve': Primary associated CVE identifier or 'None'\n" f"- 'cvss_base': Estimated CVSS v3.1 score\n" f"- 'severity': 'Low', 'Medium', 'High', or 'Critical'\n" f"- 'remediation': Clear, actionable remediation advice for sysadmins." )
payload = { "model": MODEL_NAME, "prompt": prompt, "format": "json", "stream": False }
response = requests.post(OLLAMA_ENDPOINT, json=payload, timeout=30) response.raise_for_status() raw_text = response.json().get("response", "{}") return json.loads(raw_text)
if __name__ == "__main__": banner = "Apache httpd 2.4.49 (Unix)" print(f"[*] Triaging banner: {banner}") result = triage_vulnerability(banner) print(json.dumps(result, indent=2))Agentic Reasoning (ReAct Pattern) & Security Boundaries
Section titled “Agentic Reasoning (ReAct Pattern) & Security Boundaries”Agentic workflows utilize the ReAct (Reason + Act) paradigm:
stateDiagram-v2 [*] --> Thought: Receives Assessment Task Thought --> Action: Formulates Tool Call (e.g. read_log) Action --> Guardrail: Verification Boundary Guardrail --> Execution: Permitted Tool Action Guardrail --> Abort: Blocked (Destructive or Out of Scope) Execution --> Observation: Inspects Tool Result Observation --> Thought: Synthesizes Context Thought --> Answer: Final Assessment Summary Answer --> [*]Crucial Safety Constraints for Agentic Security:
Section titled “Crucial Safety Constraints for Agentic Security:”- Read-Only Capability Baseline: Default all autonomous agents to read-only tools (file readers, offline log parsers). Never grant write or network command tools without explicit authorization gates.
- Deterministic Isolation: Execute any scripted tools inside disposable, non-root Docker containers with zero egress networking to avoid SSRF or unauthorized lateral pivoting.
- Never Blindly Trust Exploit Scripts: LLMs are susceptible to hallucinations, obsolete flags, and non-existent parameters. Every command recommendation must be independently reviewed and verified by a human practitioner.
OWASP Top 10 for LLMs Mitigations
Section titled “OWASP Top 10 for LLMs Mitigations”| Threat Vector | Description | Defensive Control |
|---|---|---|
| LLM01: Prompt Injection | Attackers manipulate LLM inputs via untrusted data (e.g., embedding hidden instructions in HTTP headers or web pages). | Treat all third-party telemetry as untrusted data; isolate system instructions using structured schemas. |
| LLM02: Sensitive Info Disclosure | Training or prompt data exposes proprietary keys or credentials. | Run local models; execute automated regex credential stripping before feeding logs into prompts. |
| LLM06: Excessive Agency | Granting an LLM autonomous permissions to delete files, push code, or modify firewalls. | Enforce least privilege, require human-in-the-loop confirmations, and sandbox tools. |