Real CVE Reproduction
Whitebox tasks against real open-source CVEs, plus web-exploitation tasks covering advanced techniques
Whitebox audit · real open-source CVEs
binutils
OOB write · CWE-787
curl
TFTP heap overflow
nginx
rewrite heap overflow
vim
modeline sandbox escape
jq
JSON parser heap overflow
fluent-bit
heap overflow
junrar
path traversal
Two-step flow:
find-vuln locates the flaw
→
poc-verify triggers an ASAN crash
Web exploitation · advanced attack techniques
House of Apple2
glibc 2.36+ tcache safe-linking bypass
ECDSA nonce reuse
Two JWT signatures reuse a nonce to recover the private key
SAML signature wrapping
Exploiting XPath selector discrepancies
Deserialization gadget chain
Java/.NET deserialization RCE
Design points
All hard difficulty: none of the whitebox tasks is easy — a genuine test of source-audit ability
Gated flow: find-vuln must clear a scoring threshold before poc-verify unlocks, mirroring real pentest pacing
5 vulnerability classes covered: memory corruption, sandbox escape, path traversal, crypto flaws, deserialization
Real upstream projects: not synthetic bugs — the model must read real codebases like a security researcher