Skip to content
ARTEX.sh

ARTEX benchmark scores:67.83 on Baidu Agent+, 89.78 on TSecBench.

ARTEX won the first Baidu BSRC Agent+ attack and defense challenge on 2026-09-03. Here are both scores, how the final was judged, and every other test of ARTEX we found.

Baidu Agent+ live range

2026-08-16

67.83/100

7th place, 5h 08m, 456M tokens

Tencent TSecBench

2026-09-03

89.78/100

6th place, 5h 03m, 271M tokens

Source: CN-SEC repost of the contest write-up

How ARTEX won Baidu Agent+:half live range, half jury.

Run by Baidu BSRC with Kuaishou, Didi, Duxiaoman and others. The press names 李浦华 (Li Puhua) as the ARTEX team lead.

teams online
150+
live range and jury
50/50
overall, national champion
1st
offline final
Chengdu

What the jury scored

  1. 01Full-process automation
  2. 02Vulnerability discovery rate
  3. 03False-positive rate
  4. 04Code-audit scale
  5. 05Time to the first high-severity bug
  6. 06LLM running cost
  7. 07Share of time spent on human checks

Every ARTEX test on record,from contest scores to fork harnesses.

  • Baidu BSRC 'Agent+' attack/defense challenge, overall

    ARTEX (v0.3.x, team led by 李浦华)

    National champion, 1st overall (50% live range + 50% jury review)

    2026-09-03cn-sec.com

  • Baidu Agent+ live range

    ARTEX

    67.83/100, 7th place; 5h08m18s effective time; 456,495,944 tokens

    2026-08-16cn-sec.com

  • Tencent TSecBench blind-test range

    ARTEX

    89.78/100, 6th on the latest leaderboard; 5h03m40s effective time; 270,797,493 tokens

    reported 2026-09-03cn-sec.com

  • BaaRai/ARTEX L3 benchmark harness (Juice Shop ground truth, 8 entries, independent LLM judge, Wilson 95% CI)

    BaaRai fork

    Harness only; results dir is gitignored, no numbers published

    2026-10-12 (PR #59)BaaRai/ARTEX

  • docker-compose.bench.yml in upstream

    ARTEX upstream / RuoJi6

    Compose stub for an artex-bench container (Dockerfile.bench, bench/.env not in repo); no results

    undatedRuoJi6/ARTEX

  • /btw side-question validation

    ARTEX feature test

    Engineering test table, all pass; not a capability benchmark

    2026-09-10RuoJi6/ARTEX

ARTEX in real attacks:speeds from threat reports.

Not benchmarks: these come from incident reports on attackers who used ARTEX.

  • 15 min

    fastest jump from a first stolen account to deep access. About 4 hours is typical

    GTIG

  • 27 in 23 s

    events while probing AWS, in bursts of 2 to 25 seconds

    GTIG

  • 2,900/min

    sign-in attempts against 2,250+ cloud tenants

    Microsoft via The Hacker News

Frameworks named next to ARTEX:PentAGI, Cairn, Hexstrike and AgentZero.

  • PentAGI

    VXControl, Go, MIT. About 14,700 GitHub stars and 4 sub-agents.

  • Cairn

    oritera/Cairn, 4,756 stars. The ARTEX README lists it as its reference.

  • Hexstrike and AgentZero

    Named by GTIG with PentAGI and CAIRN as agentic frameworks attackers also abused.

  • ARTEX Pro

    ThreatBook's commercial edition on RedAgents, sold next to Cyber (YanJi), Yanxun and CyberStrikeAI.

Follow the ARTEX story.

New forks, releases and reports, dated and sourced.

Read the timelineHow ARTEX works
Isometric plane with glowing light columns