We emulate
the AI-era attacker.

Our agent found a permissions flaw in a test web app. Before touching any client data it stopped and asked a person for approval. Result: one High finding, validated, in 5 min 28 s.

1. Recon started
2. Found a sensitive user
3. Abused a logic flaw to reach restricted test data
4. Asked a person for approval
5. Documented the attack chain
6. Showed the leak with made-up data
7. Wrote the report
8. Validated it, no false positives
9. Gave defenders clues to hunt it

tr_4c81e07d2a9fweb-recon-agentneeds review5m 28s31,406 → 5,221 tok

1. agentRuns the engagement invoke_agent web-recon-agentroot span, starts at 0 ms, 5m 28sscope: app.example.com (203.0.113.24)warn
2. toolScans the test app http_probechild of invoke_agent web-recon-agent, starts at 412 ms, 6.85 s14 hosts, 3 web services, 61 API routesT1595.002
3. toolFinds a user who holds sensitive data account_reviewchild of invoke_agent web-recon-agent, starts at 11.3 s, 9.14 sseeded test user A: billing admin, 2 client invoicesT1087
4. toolUses a logic flaw to read A's data as B authz_diffchild of invoke_agent web-recon-agent, starts at 21.8 s, 48.2 sseeded test users: B's token returns A's test invoiceT1190warn
5. eventAsks a person before touching client data operator_checkpointchild of invoke_agent web-recon-agent, starts at 1m 12s, 2m 12soperator approves reading 2 client records
6. toolRecords the attack path with evidence evidence_capturechild of invoke_agent web-recon-agent, starts at 3m 24s, 9.47 s2 client records read; hashes logged; 4-step chainT1213
7. toolShows the leak with made-up records synthetic_exfilchild of invoke_agent web-recon-agent, starts at 3m 34s, 38.2 sschema: 14 fields; 200 synthetic invoices pulledT1567
8. llmWrites the report chat report_draftchild of invoke_agent web-recon-agent, starts at 4m 13s, 41.7 sfinding, impact, fix; 1,240 words9,318 → 1,604 tok
9. evalRe-checks it to rule out a false positive finding_scorerchild of invoke_agent web-recon-agent, starts at 4m 55s, 27.4 sHigh: broken object-level authorization; reproduced twice
10. toolHands defenders clues to hunt with ioc_exportchild of invoke_agent web-recon-agent, starts at 5m 23s, 4.11 s6 IOCs: 3 request paths, 2 user agents, 1 test account

In a lab, our agent found an unknown flaw in a network device's software. A person approved a harmless proof. The logs caught it; the endpoint security tool did not. Result: a detection gap filed in 29 min.

1. Took the device software apart
2. Found the flaw
3. Asked a person for approval
4. Ran a harmless proof
5. Checked the logs
6. Checked the endpoint security tool
7. Scored detection coverage
8. Filed the gap

tr_b19e5530c6d4purple-loop-agentgap found29m 03s41,368 → 6,209 tok

1. agentRuns the exercise in a lab invoke_agent purple-loop-agentroot span, starts at 0 ms, 29m 03slab only: edge-gw fw 4.2.1, latest releasewarn
2. toolTakes the device software apart ghidra.decompilechild of invoke_agent purple-loop-agent, starts at 1.2 s, 3m 07s1,412 functions; upload parser flagged
3. llmSpots the flaw chat analyze_diffchild of invoke_agent purple-loop-agent, starts at 3m 08s, 21.5 sunchecked length (LAB-0412); unreported, disclosure opened11,872 → 1,406 tok
4. eventAsks a person for approval operator_checkpointchild of invoke_agent purple-loop-agent, starts at 3m 30s, 6m 42sbenign marker proof approved, lab target
5. toolProves it with a harmless marker proof_runchild of invoke_agent purple-loop-agent, starts at 10m 12s, 14.9 smarker file written in the lab sandboxT1190
6. toolChecks what the logs saw elastic.esqlchild of invoke_agent purple-loop-agent, starts at 10m 29s, 3.27 slogs-edge-gw* /upload*: 4 hits
7. toolChecks what the endpoint security tool saw falcon.event_searchchild of invoke_agent purple-loop-agent, starts at 10m 33s, 8.84 schild process on host; 0 detectionsT1059warn
8. evalScores how much was caught detection_coveragechild of invoke_agent purple-loop-agent, starts at 10m 59s, 1.08 spartial: 1 of 3 stages detectedwarn
9. toolFiles the gap for the defenders tracker.file_gapchild of invoke_agent purple-loop-agent, starts at 11m 01s, 2.73 sDET-0412 opened; 2 detection rules proposed