gnosis retweeted
Another result for Out of Bounds (@oobs_io) at #Pwn2Own!
ByungYoung Yi (@yibarrack) and KeunHo Kim (@gn0sis__) exploited the Philips Hue Bridge Pro using 5 bugs, earning $12K and 2.5 Master of Pwn points.
Two entries down, one more to go.
It seems to be a day for collisions. ByungYoung Yi (@yibarrack), and KeunHo Kim (@gn0sis__) of Out of Bounds (@oobs_io) used 5 bugs to exploit the Phillips Hue Bridge Pro, but 4 of those bugs were previously known. They still earn $12,000 and 2.5 Master of Pwn points. #Pwn2Own
So excited to have pulled off an exploit at my first Pwn2Own! Let's goooo!
#Pwn2Own
It seems to be a day for collisions. ByungYoung Yi (@yibarrack), and KeunHo Kim (@gn0sis__) of Out of Bounds (@oobs_io) used 5 bugs to exploit the Phillips Hue Bridge Pro, but 4 of those bugs were previously known. They still earn $12,000 and 2.5 Master of Pwn points. #Pwn2Own
1
7
278
Excited to be at Pwn2Own Ireland for the first time! Good luck, everyone! If you’re here too, feel free to reach out!
#Pwn2Own
We’re at Pwn2Own again!
Out of Bounds is bringing 3 entries to Pwn2Own Ireland 2026.
Full schedule:
zerodayinitiative.com/blog/2…
If you’re in Cork, hit me up!
And stay tuned for updates 👀
1
188
My first paper is now published in IEEE Access.
Starting with an SCIE journal felt overwhelming, but I’m grateful we brought it to completion.
This work taught me a lot: careful problem definition, honest limitations, and verifiable claims.
ieeexplore.ieee.org/document…
133
SCOUT v2.8.0.
AutoPoC now uses Exploit Pattern RAG.
→ metadata-backed refs
→ Web/Config/IPC matching
→ sink + trigger scoring
→ adaptation-first prompting
→ contamination guard for endpoints/IPs/products
Not copy-paste. Adapt the exploit pattern.
github.com/R00T-Kim/SCOUT
63
SCOUT v2.7.2.
Rebuilt detection in 4 weeks. LATTE + LARA + sink 28→51+.
Recall 0.143 → 0.143.
Hypothesis killed. Pivoted to compliance/audit evidence engine.
CRA / FDA 524B / ISO 21434 / UN R155 mappings shipped.
github.com/R00T-Kim/SCOUT
55
SCOUT v2.6.1.
0.7 from symbol co-occurrence and 0.7 from P-code verification were the same number. Split into 3 axes:
→ evidence_tier (6 levels)
→ confidence (within-tier)
→ priority_score
SBOM silent bug: R7000 0 → 4.
Corpus 1,123: 1,110 / 4 / 9.
github.com/R00T-Kim/SCOUT
47
"All your findings are LOW quality." So we rebuilt the engine.
SCOUT v2.4.0: → Ghidra P-code taint (source→sink) → 3-tier confidence (0.40/0.55/0.75) → +4 rules (SQLi, FSB, path traversal, SSRF)
RT-AX88U: 15 findings, max conf 0.60
github.com/R00T-Kim/SCOUT/re…
45
SCOUT v2.2.0 is out.
Tier 1 benchmark results on 1,123 FirmAE firmware images across 8 vendors:
99.2% analysis success rate
3,523 findings
13,893 CVE matches
42-stage pipeline fully executed
Next: Tier 2 with LLM integration and better FP verification.
github.com/R00T-Kim/SCOUT
72
Drop a firmware blob. Get SARIF findings, CycloneDX SBOM, and a hash-anchored evidence chain.
SCOUT: 34-stage pipeline, zero pip deps, pure Python.
Static-only findings hard-capped at 0.60 — we don't inflate confidence.
Sample results included.
github.com/R00T-Kim/SCOUT
54






