v3.1: fix residual false positives on short bare commands (pwd, whoami, echo X, cat X) — benign-command acc 0.93->0.98, realistic metrics unchanged (macro-F1 0.931, malicious recall 0.945)

README.md

@@ -45,10 +45,12 @@ form. This reflects real-world use, unlike an in-distribution test split.
 | benign — questions | 0.96 |
 | malicious recall | 0.945 |
 
-On a separate **held-out shell-command set** (ordinary developer commands like
-`git status`, `npm install`, `docker ps -a`, plus real command-style attacks):
-benign-command accuracy **0.93**, attack recall **1.0**. Earlier versions wrongly
-flagged plain commands as attacks; this is fixed.
+On a separate **shell-command eval** (developer commands from `git status` and
+`npm install` down to trivial `pwd`, `whoami`, `cat file.txt`, `echo 'Hello'`,
+plus real command-style attacks): benign-command accuracy **0.98**, attack recall
+**1.0**. Earlier versions wrongly flagged plain shell commands as attacks — a
+problem for tool/agent guardrails that screen commands before running them; this
+is fixed.
 
 Threshold is 0.5 (`argmax`). Raise it for fewer false positives, lower it to
 catch more attacks.

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:af60710e620ab20f202f954246f71d7c46c33239b872bd0586314be6ce919632
+oid sha256:b3aec2a5d19b369d9bb92020cc296c805e6c872dc31a988e2428c50b21d9463f
 size 53330319