1. Ostorlab Neutron reaches 97.1% on CyberGym.
   On screen: The verified solve rate fills the screen.

2. 1,463 tasks produced verified proofs of concept.
   On screen: Exactly 1,463 of 1,507 task cells fill.

3. Neutron achieved 100% detection across CyberGym’s 1,507 tasks.
   On screen: Detection and localization of known benchmark vulnerabilities.

4. CyberGym includes software you already know.
   On screen: Python: f-string tokenizer off-by-one. NGINX: invalid HTTP/2 stream identifiers. FFmpeg: H.264 out-of-bounds table access. Wireshark: packet-parser length underflow.

5. curl’s wildcard parser could read past an unfinished pattern.
   On screen: The actual pattern *[][ is shown as bytes. A read ring reaches NUL, then moves beyond the string.

6. The fix checks the string terminator before every state.
   On screen: Show the upstream patch: c = **p; if (!c) return SETCHARSET_FAIL; before switch(state).

7. For FAAD2, Neutron reconstructed the input needed to reach the decoder.
   On screen: Raw source, the 29-byte harness preamble, and the decoder appear in sequence.

8. FAAD2 used channel mappings without rechecking the channel limit.
   On screen: Show the pce.channels > MAX_CHANNELS guard before internal_channel indexing, as added by the upstream patch.

9. The input declared 93 channels against a 64-entry buffer.
   On screen: 64 available entries in blue; 29 excess channels in red.

10. A 134-byte proof reproduced the overflow.
   On screen: The same FAAD2 input triggers a stack-buffer overflow before the patch and exits cleanly after it.

11. Average estimated inference cost was $1.04 per task.
   On screen: The cost is labelled per benchmark task.

12. Turn vulnerability findings into executable proof.
   On screen: The outcome line settles, followed by the Ostorlab logo.