SwishOS Developer Documentation

Target-agnostic HTTP red-teaming harness, deterministic shape-based detection, and tool-call sequence analysis for production AI agents.

Package Installation & CLI Harness

Python SDK Runtime Integration

from agentic_redteam.patching.guardrails import check_ingress, sanitize_egress

# Ingress: refuse the request before the agent ever sees it
blocked, rule_id, reason = check_ingress(user_input)
if blocked:
    raise SecurityError(f"{rule_id}: {reason}")

# Egress: scrub API keys and PII out of the response
safe_output, redactions = sanitize_egress(agent_response)

Deterministic Runtime Sandbox

Generates hardened container configuration for running agents under gVisor (runsc) user-space kernel isolation, with read-only root, cgroup memory caps, and iptables rules dropping cloud metadata egress.

service = generate_gvisor_docker_compose_service(
    image="your-agent:latest",
    memory_limit="256m",
)
# runtime: runsc  |  read_only: true  |  cap_drop: ALL

Deterministic Response Analysis

Inspects responses for the shape of a leak -- validated card numbers, IBANs, cloud metadata, dangerous code -- rather than asking a model whether the answer looked unsafe. The same input always produces the same finding.

hits = find_pii(response)          # Luhn / mod-97 validated
if reveals_cloud_metadata(response):
    flag("SSRF: cloud metadata reached")

Response Body Scrubbing

Strips provider API keys and PII from agent responses before they leave the process, covering current key formats such as Stripe, AWS, and GitHub tokens.

safe, redactions = sanitize_egress(response_body)
if redactions:
    log.warning("redacted %d secret(s) from egress", len(redactions))