Claude Code with Fable 5.1 posts 87.2% FuncPass and 37.4% SecPass — a new #1 on our board — while running as fast as Fable 5, costing a third of Opus 5, finishing every task with zero timeouts, and drawing less than half the confirmed cheating of both predecessors. | Blog | Endor Labs

What it is

An Endor Labs "Agent Security League" benchmark post scoring Claude Code + Fable 5.1 against Opus 5 and Fable 5.0. The summary already carries the leaderboard numbers, the methodology framing, and the three CVE identifiers. What it does not carry is the fix invariant behind each of those three "hall of fame" cases and the memorization-adjustment detail — the parts an engineer could actually reuse.

The three fix patterns

These are historical, already-patched flaws; the value is the concrete secure pattern each demonstrates.

• Home Assistant — CVE-2018-21019 / CWE-200. The error log was served via register_static_path, and static routes bypass the view auth middleware entirely — anyone on the network could GET /api/error_log. Fix: serve it through an authenticated HomeAssistantView instead of a static path (patch similarity to golden: 0.46).

• html-sanitizer — CVE-2024-34078 / CWE-79. NFKC normalization ran after sanitization, folding full-width characters (e.g. ＜) into ASCII < — so <script> encoded in full-width chars passed sanitization as text, then folded back into live markup. Fix: normalize before sanitizing. Fable 5.1 additionally stripped control chars / lone surrogates (patch similarity 0.09).

• Vyper — CVE-2023-31146 / CWE-787. In dynarray codegen, self-referencing ops like a = [a[0], a[1], a[2]] stored the length word before copying data, so bounds checks passed against a not-yet-populated array — out-of-bounds read/write. Fix is an ordering invariant: copy the data first, then store the length word (patch similarity 0.27).

Methodology worth noting

Scores are published with memorization removed (confirmed training-recall solves stripped). Fable 5.1 loses only 6.7 FuncPass / 3.9 SecPass points on that adjustment, roughly a third of what Opus 5 and Fable 5.0 lose — so its lead survives correction. Confirmed cheats: 17 (Fable 5.1) vs 38 (Opus 5) vs 40 (Fable 5.0), and 12 of 17 are "recall-then-diverge."

To catch that last pattern, Endor now grades the edit trajectory — first/peak security overlap, suddenness of change, and string chronology — because early recall stays visible even when the final patch has diverged. A small set of "overly strict" instances demanding unguessable implementation details are excluded from the denominator but kept as tripwires: a pass triggers trajectory inspection rather than automatic credit.

Caveats

Harness versions differ (Fable 5.1 on Claude Code 2.1.258; the others on 2.1.156), so the model is not the sole variable. Fable 5.0's ~$417 cost is a lower bound — 16 aborted runs left no recoverable spend record. Absolute SecPass remains low (37.4%); the author notes progress on secure coding "has been slow."