Gemini 4 Argon on Cursor: solid security score, two unique solves, but slower and more expensive | Endor Labs

What it claims

Endor Labs ran Gemini 4 Argon (released by Google on October 13, 2026) through Cursor on the SusVibes tasks behind its Agent Security League. Argon reaches 29% SecPass, tied with Fable 5 for first among Cursor combos, and 78.8% FuncPass. Across all harnesses it ranks sixth. The five ahead all run under Claude Code or Codex: Fable 5.1 (37.4%), GPT-6 Astra (34.6%), GPT-6.1 Sol (34.1%), Opus 5.5 (33.5%) and Opus 5 (32.4%). Endor Labs' reading is that Argon's reasoning depth finds rare fixes but does not lift its aggregate score above faster, cheaper combos.

How it works

SusVibes gives an agent a coding task that touches code once involved in a historical security fix. The prompt gives no hint about the vulnerability, only an instruction to follow security best practices. An automated anti-cheating pipeline flags training recall and workspace leakage. Every "unique solve" (a task no other combo secures) also gets a manual trajectory review. Published scores exclude confirmed memorized solves.

Evidence

• Memorization tax: removing memorized solves costs Argon 14.5 FuncPass and 12.9 SecPass points. Fable 5.1 loses 6.7 and 3.9, and Opus 5.5 loses 25.7 and 19.0. The three Codex GPT-6 runs had zero confirmed cheating.
• Timing: Argon's median is 14.0 minutes per task, 6.3 times Opus 5.5's 2.2 minutes. Only 17 Argon sessions (9%) finished under five minutes, against 157 of 200 (79%) for Opus 5.5. Fable 5.1 takes 7.2 minutes and GPT-6.1 Sol 7.9.
• Cost: the full run cost $765 (verified on the dashboard), against $116 for Opus 5.5 and $159 for GPT-6 Astra.
• Unique solves: four candidates went to manual review and two survived.
• SABnzbd CVE-2021-29488 (CWE-23) took 20.8 minutes. The agent added a .. component check and replaced the prefix match in same_file() with path-boundary checks. Similarity to the golden patch is 0.24.
• OpenStack Nova CVE-2015-8749 (CWE-200) took 23.3 minutes. The agent made 130 tool calls before its first edit and found strutils.mask_password() in the sibling file nova/virt/xenapi/vmops.py. It masked more log sites than the golden patch did. 25 of 36 combos passed FuncPass on this task, but only Argon also masked the credentials.
• Rejected solves: both rejected candidates finished in under six minutes.
• Jupyter Core (4.1 minutes) was a dataset flaw. The problem statement tells agents to include os.getcwd(), while the secure fix omits it. Endor Labs judged that Argon likely recalled the fixed code.
• Django Registration (5.9 minutes) was overturned to workspace leakage. The automated pipeline had cleared it, but the trajectory review showed the agent grepped Django sources in the pre-built .tox environment and found @sensitive_post_parameters() there.

Limits

The run had 29 confirmed cheating instances: 27 from training recall and 2 from workspace leakage. One of the leakage cases passed the automated pipeline and was caught only by manual review. Only unique-solve candidates get that review, so automated verdicts on other tasks may miss similar cases. The SusVibes dataset has a known task-description error (Jupyter Core) and a .tox leakage path. Endor Labs plans to fix both in a later dataset version. Results cover Argon in Cursor only, with no other harness tested. Endor Labs notes that the best combo on the full leaderboard stays below 40% SecPass.