Blog · Claude Fable 5.2

Claude Fable 5.2: new model, same holes

Claude Fable 5.2 is Anthropic’s newest coding model. It writes less exploitable code than Opus 5.5, but it still leaves the security holes that are already in your app.

LeoTrace12 November 2026

We wanted to know whether a newer model means safer code, so we ran Fable 5.2 and Opus 5.5 through two of our security benchmarks. Both ran in Claude Code on the same tasks, three times each.

Share of runs left vulnerable, lower is better. On 33 apps that already had a security hole, the hole was still there at the end in 84% of runs for Claude Opus 5.5 on its own, 94% for Claude Fable 5.2 on its own and 29% for Fable 5.2 with LeoPrevent. On 64 real CVE tasks, the original exploit still worked in 38%, 19% and 9% of runs. FLAWED APPS · HOLE STILL THERE AT THE END Opus 5.5 alone 84% Fable 5.2 alone 94% Fable 5.2 + LeoPrevent 29% REAL CVES · CODE STILL EXPLOITABLE Opus 5.5 alone 38% Fable 5.2 alone 19% Fable 5.2 + LeoPrevent 9% 0% 25% 50% 75% 100%
Share of runs left vulnerable, lower is better. Same tasks and runs with and without LeoPrevent. Full results and method.

Real CVEs: half the exploitable code

A.S.E gives the agent a real open-source project with the function behind a published vulnerability cut out, and asks it to write that function back. Then it runs the original exploit against the result.

The exploit got through Opus 5.5’s code in 38% of runs and through Fable 5.2’s in 19%. That is a big improvement, but it still means one run in five put a known vulnerability back into the project.

Flawed apps: the hole stays in

Our other benchmark uses 33 small apps that already have a security hole in them, as most real codebases do, and asks the agent for a feature in the same part of the code.

Fable 5.2 finished with the hole still in the app in 93 of 99 runs. Opus 5.5 did that 83 times. On this one the newer model was slightly worse.

The task didn’t ask either model to fix the hole, and coding agents tend to stick to the task. Existing security debt rarely makes it into a ticket, so it stays where it is.

With LeoPrevent

LeoPrevent reviews the agent’s work before the agent finishes, including code that was already there, and tells the agent what it found. The agent then fixes it in the same turn.

With LeoPrevent, Fable 5.2 left the hole in 29 apps instead of 93, about a third as many. On real CVEs its exploitable code dropped from 19% to 9%.

Moving from Opus 5.5 to Fable 5.2 helped on one benchmark and hurt on the other. LeoPrevent at least halved the vulnerable results for both models on both benchmarks, and with it the two models ended up close to each other: 29 and 23 apps, 9% and 13% of CVE runs.

So a newer model helps, but it doesn’t clean up the security debt that is already in your code. For that, the agent’s work needs checking before it ships.

Try LeoPrevent with your agent

The method and the verdict of every run are on the results page.