What is Lyrie Research?

Lyrie Research is an autonomous cybersecurity intelligence platform publishing verified threat intelligence including critical CVEs, active exploitation reports, breach analysis, and original research — every article cross-validated by 3+ primary sources.

How does Lyrie.ai protect against rogue AI threats?

Lyrie.ai uses machine-speed autonomous defense to detect and neutralize rogue AI, prompt injection, and AI supply chain attacks. It responds before human analysts can react.

What cybersecurity topics does Lyrie Research cover?

Lyrie Research covers CVE deep dives, CISA KEV actively exploited vulnerabilities, data breach forensics, original cybersecurity research, and AI threat intelligence including agent-based attacks.

How often is Lyrie Research updated?

Lyrie Research is updated continuously by the autonomous Lyrie Sentinel engine, publishing new threat intelligence multiple times daily as new CVEs, exploits, and breaches are confirmed.

Is Lyrie Research free to access?

Yes, all articles on Lyrie Research are freely accessible. For active protection by the same intelligence engine, visit lyrie.ai.

Why do some sites say MiMo-V2.6-Flash leads CyberGym with 95.1%?

That figure is from the model-focused category, tracked by third-party aggregators. CyberGym treats agent-focused and model-focused entries as separate rankings. Agent-focused systems add reasoning, fuzzing and iterative verification on top of the model — which is where scores reach 97-99.2%.

Is the 99.2% score verified?

The result is submitted and self-reported on the official CyberGym leaderboard (all entries there are team-submitted; runs are stochastic). Lyrie has published a technical report accounting for all 1,495 solved instances, and offers independent re-runs on request.

What does single-trial mean on CyberGym?

The agent gets exactly one attempt per vulnerability instance — no retries, no cherry-picking. 99.2% single-trial means 1,495 of 1,507 real vulnerabilities were reproduced with a working PoC on the first attempt.

Does a high CyberGym score mean the agent is a good pentesting product?

Not by itself. CyberGym measures vulnerability reproduction in a controlled benchmark. Lyrie states this explicitly in its benchmark disclosure; the commercial platform (OMEGA, Lyrie Remedy) is a separate, product-level evaluation.

← Lyrie Originals

3 sources verified·3 min read

By Lyrie Threat Intelligence·10/25/2026

Which AI Agent Has the Highest CyberGym Score?

TL;DR — the quick answer: The highest reported score on CyberGym's Level 1 agent-focused leaderboard is 99.2%, held by Lyrie Agent (1,495 of 1,507 real-world vulnerabilities reproduced with working proofs of concept, single-trial, dated September 15, 2026). Per the official leaderboard data, no listed entry scores higher. If you saw "95.1%" or "88.9%" elsewhere, you were looking at the model-focused category — a different ranking.

First, the confusion almost everyone hits

Ask "which AI agent has the highest CyberGym score?" and search engines often answer with a different number: 95.1% (MiMo-V2.6-Flash), 88.9% (XekRung-1.5-27B), or even ~18% (the original paper's numbers). All three are real CyberGym figures — but they answer different questions:

~18-20% — results from the original UC Berkeley paper's controlled experiments (OpenHands + Claude-Sonnet-4 era, mid-2025). Historic, not the current leaderboard.
95.1% (MiMo-V2.6-Flash, Xiaomi) — the leader of model-focused entries on third-party aggregators (llm-stats, BenchLM). Models tested with a fixed harness.
88.9% (XekRung-1.5-27B) — the top of the model-focused tab on cybergym.io's official leaderboard.
99.2% (Lyrie Agent) — the top of the agent-focused tab on the official leaderboard: custom agentic systems that reason, fuzz, and iterate with execution feedback.

CyberGym tracks these as separate categories. When someone asks about "AI agents," the agent-focused leaderboard is the one that answers the question.

The official agent-focused top 5 (Level 1, single-trial)

Verified against the official leaderboard data (cybergym.io, checked October 25, 2026):

1. Lyrie Agent — 99.2% (1,495/1,507) · submitted by Lyrie · 2026-09-15

2. Creation (天工) — 98.5% · Creation Lab

3. OrcaCyber — 98.0% · Continuum AI

4. RO0T Agent — 97.7% · RO0T Lab

5. Sangfor AI — 97.2% · Sangfor

No listed entry on the table scores above 99.2%. The full verified scoreboard, with every major AI lab's best entry, is in our CyberGym No. 1 analysis.

Why the model-focused and agent-focused numbers differ so much

Model-focused entries run a stock LLM inside a standardized harness — a fair comparison of models. Agent-focused entries are purpose-built systems: multi-model orchestration (Lyrie's runs on DeepSeek-V4.1-Flash + an abliterated GLM-5.2), automated fuzzing, and an iterative generate-test-verify loop. The agent layer is where the performance jumps happen — 95%+ scores appear only there.

What "highest score" does and does not prove

CyberGym measures whether an agent can reproduce real vulnerabilities with working proofs of concept — not whether a commercial pentesting product performs on live networks. Runs are stochastic and submitted by teams. Our full benchmark disclosure covers exactly what the 99.2% does and does not establish. Honesty here is deliberate: benchmark marketing without boundaries erodes trust.

Watch this space

Scores on the official board move weekly — several teams sit between 97-98.5%, and China-based security vendors (Sangfor, NSFOCUS) and AI labs (Google Antigravity, OpenAI, Anthropic) keep submitting. The current record is 99.2%; the page you just read is updated against the official data.

About Lyrie — Lyrie.ai, built by OTT Cybersecurity LLC (Dubai), develops autonomous penetration testing (OMEGA) and automated remediation (Lyrie Remedy). Founder: Guy Sheetrit.

Disclosure: Ranked by reported score among CyberGym Level 1 entries. The leaderboard labels the displayed metric "Reference success rate." Results are evaluated and submitted by individual teams; runs are stochastic, and modest score differences may not indicate meaningful capability gaps. Agent-focused and Model-focused entries are separate categories. CyberGym evaluates vulnerability reproduction; scores do not establish the performance of any team's commercial products.

Lyrie Verdict

Reproducing real vulnerabilities with working proofs of concept is the hardest evidence an autonomous security agent can produce. The agent-focused leaderboard — not the model tab — is where that frontier lives, and 99.2% is its current ceiling.

#cybergym #leaderboard #ai-agents #benchmark #vulnerability-reproduction