Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when reverse-engineering a binary or firmware — static triage + decompilation (Ghidra/IDA/Binary Ninja), dynamic instrumentation (GDB/Frida 17/angr), anti-reversing & packer bypass, OLLVM/VM deobfuscation, UEFI/BIOS RE & Secure Boot research, patch-diffing for n-days
.claude/skills/hypnguyen1209-reverse-engineering/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 307% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 71% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 155% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 93% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 119% | 0% |
| Technique | ATT&CK | CWE | Reference | Script | |-----------|--------|-----|-----------|--------| | Binary triage (format/arch/mitigations/strings/imports) | T1592.002, T1518.001 | CWE-1395 | references/static-triage-decompilation.md | scripts/triage.py | | Headless decompilation (Ghidra 11.4 / r2 / IDA) | T1592.002 | CWE-noinfo | references/static-triage-decompilation.md | scripts/triage.py | | AI/MCP-assisted RE (Sidekick / GhidrAssistMCP / LLM4Decompile) | T1592.002 | CWE-noinfo | references/static-triage-decompilation.md | - | | Dynamic debugging (GDB/GEF, x64dbg, conditional bps) | T1622 | CWE-noinfo | references/dynamic-instrumentation.md | - | | Frida 17 instrumentation + SSL-pin/JNI hooking | T1622, T1562.001 | CWE-noinfo | references/dynamic-instrumentation.md | scripts/frida_universal.js | | Symbolic / concolic execution (angr, Triton) | T1480.001 | CWE-noinfo | references/dynamic-instrumentation.md | scripts/deflatten_triton.py | | Anti-debug detection & bypass (PEB/ptrace/HW-bp/timing) | T1622, T1497.001 | CWE-noinfo | references/anti-reversing-bypass.md | scripts/antidebug_unhook.py | | Anti-VM / sandbox-evasion neutralization | T1497, T1497.003 | CWE-noinfo | references/anti-reversing-bypass.md | scripts/antidebug_unhook.py | | Packer unpack → OEP dump (UPX/runtime packers) | T1027.002, T1620 | CWE-noinfo | references/anti-reversing-bypass.md | scripts/antidebug_unhook.py | | String / API-hash deobfuscation (Unicorn emulation) | T1027.013, T1140, T1027.007 | CWE-noinfo | references/deobfuscation.md | scripts/string_decrypt_emu.py | | OLLVM control-flow-flattening de-flattening | T1027.009, T1027 | CWE-noinfo | references/deobfuscation.md | scripts/deflatten_triton.py | | VM-protector devirtualization (VMProtect 3.x / Themida) | T1027.009 | CWE-noinfo | references/deobfuscation.md | scripts/deflatten_triton.py | | Firmware extraction (binwalk/squashfs/QEMU emulation) | T1542.001 | CWE-1263 | references/firmware-uefi.md | scripts/uefi_triage.py | | UEFI/BIOS RE + Secure Boot bypass research | T1542.001, T1542.003 | CWE-347 | references/firmware-uefi.md | scripts/uefi_triage.py | | Patch diffing → n-day root cause (BinDiff/Diaphora/ghidriff) | T1203, T1592.002 | CWE-noinfo | references/patch-diffing-protocol.md | scripts/patchdiff_fetch.py | | Protocol / file-format inference (Netzob/Kaitai) | T1592.002 | CWE-noinfo | references/patch-diffing-protocol.md | scripts/proto_infer.py |
bash# 0. Triage: format, arch, mitigations, packer entropy, strings, imports, capabilities → JSON python3 scripts/triage.py ./sample -o out/triage.json # 1. Headless decompile to C (Ghidra 11.4 analyzeHeadless wrapper inside triage.py --decompile) python3 scripts/triage.py ./sample --decompile --ghidra "$GHIDRA_HOME" -o out/ # 2. If packed/protected → defeat anti-debug, dump OEP (run under x64dbg+ScyllaHide on Windows) python3 scripts/antidebug_unhook.py --scan ./sample # enumerate anti-debug primitives first # 3. Dynamic: attach Frida (Android/Linux/Win), universal SSL-unpin + native trace frida -U -f com.target.app -l scripts/frida_universal.js --no-pause # 4. Deobfuscate: emulate string decryptor over all xrefs; de-flatten OLLVM with Triton python3 scripts/string_decrypt_emu.py ./sample --func 0x401500 --auto-xref -o out/strings.txt python3 scripts/deflatten_triton.py ./sample --func 0x401abc -o out/cfg.dot # 5. Firmware: carve + identify + map UEFI DXE/PEI attack surface, scan for PKfail/known hashes python3 scripts/uefi_triage.py firmware.bin -o out/fw/ # 6. n-day: fetch pre/post-patch Windows binary from winbindex and diff python3 scripts/patchdiff_fetch.py --pe afd.sys --kb-after KB5050000 -o out/diff/ # 7. Unknown protocol: infer fields/state machine from a pcap, emit Kaitai + Wireshark dissector python3 scripts/proto_infer.py capture.pcap --port 4444 -o out/proto/
| Technique | Telemetry / IOC | Detection (Sigma/EDR) | OPSEC note | |-----------|-----------------|------------------------|------------| | Static triage / decompile | None (offline on analyst box) | N/A — runs in lab | Analyze copies in an isolated, snapshotted VM; never on the target | | Frida / dynamic hooking | frida-agent.so in /proc/self/maps, port 27042, gum-js-loop/gmain threads, ptrace on target | App-side RASP (Talsec/DeepID), frida-string scans, EDR userland hook tripwires | Rename agent, use gadget+-l script mode, embed gadget in APK to dodge port checks | | Anti-debug bypass | Debug registers DR0-7 set, PEB.BeingDebugged flips, hooked Nt prologues | Self-integrity checks, KiUserExceptionDispatcher checks, TitanHide-vs-malware arms race | Prefer kernel TitanHide/HyperHide for hardened packers; snapshot before each run | | Packer/OEP dump | New RWX region, IAT rebuild, written dump.exe on disk | EDR RWX-alloc + tail-jump heuristics (lab only) | All in lab; dumped sample is for analysis, not redeployment | | String/API-hash decrypt (Unicorn) | None on target (offline emulation) | N/A | Pure offline; safe — no sample execution of network/FS code | | Firmware / SPI flash dump | Hardware: chip-clip on SPI; software: chipsec/flashrom reads | Boot Guard / measured boot (TPM PCRs), Binarly/CHIPSEC verifiers | Physical/authorized only; flashing back a modded image is destructive & loud | | Secure Boot bypass research | New unsigned bootloader in ESP, MokList/dbx anomalies, unexpected bootmgfw hash | Measured boot PCR0/2/4/7] drift, dbx revocation checks, ESET/Binarly scanners | Lab VMs / disposable hardware; document, never persist a real implant | | Patch diffing | None on target (fetches public binaries) | N/A | Public Microsoft/distro binaries; n-day PoC stays in lab until disclosure/ROE | | Protocol inference | Replays captured/crafted packets at the service | IDS on malformed/replayed frames; rate anomalies | Throttle active probes; prefer passive pcap when a live target is in scope |
analyzeHeadless + PyGhidra, r2/rizin & IDA decompile flows, and the 2025 AI/MCP-assisted layer (Binary Ninja Sidekick, GhidrAssistMCP, LLM4Decompile / SK²Decompile) with verification discipline.trace_proof artifact. Backed by scripts/rr_root_cause.sh (degrades gracefully where rr is absent). Runs in .devcontainer/.--coverage, run the witness, assert the line is not #####) and -finstrument-functions/rr/uftrace function traces; the coverage_proof/trace_proof the native-bug bar in validate_findings.py requires before [CONFIRMED].scripts/cve_diff.py — multi-source canonical fix-commit discovery (OSV/NVD/GitHub Advisories) with de-dup by (repo, sha) + OSV GIT-range fixed events, then a scope-gated, git_safe-hardened clone + git diff fix^..fix. Use it to locate the patch before BinDiff/ghidriff; chain into /engage.crash for root-cause + exploitability.| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 12,240 | 29,274 | +139% | 1 | 1 | 0% | 2,272 | 9,246 | +307% | 0 | 0 | — |
case-07 | fail→pass | 29,509 | 3,570 | -88% | 1 | 1 | 0% | 2,149 | 3,676 | +71% | 0 | 0 | — |
case-13 | pass→pass | 18,985 | 15,383 | -19% | 1 | 1 | 0% | 2,880 | 5,467 | +90% | 0 | 0 | — |
case-02 | fail→fail | 19,868 | 22,169 | +12% | 1 | 1 | 0% | 3,929 | 7,421 | +89% | 0 | 0 | — |
case-03 | fail→fail | 18,623 | 15,438 | -17% | 1 | 1 | 0% | 3,467 | 6,135 | +77% | 0 | 0 | — |
case-04 | fail→pass | 8,777 | 3,620 | -59% | 1 | 1 | 0% | 1,424 | 3,627 | +155% | 0 | 0 | — |
case-05 | fail→pass | 24,774 | 12,596 | -49% | 1 | 1 | 0% | 1,969 | 3,793 | +93% | 0 | 0 | — |
case-06 | fail→pass | 30,138 | 4,614 | -85% | 1 | 1 | 0% | 1,745 | 3,813 | +119% | 0 | 0 | — |
case-08 | fail→pass | 12,564 | 3,247 | -74% | 1 | 1 | 0% | 1,856 | 3,601 | +94% | 0 | 0 | — |
case-09 | fail→pass | 10,281 | 13,325 | +30% | 1 | 1 | 0% | 1,719 | 4,049 | +136% | 0 | 0 | — |
case-10 | fail→pass | 43,216 | 24,570 | -43% | 1 | 1 | 0% | 1,951 | 5,970 | +206% | 0 | 0 | — |
case-11 | fail→pass | 24,747 | 13,200 | -47% | 1 | 1 | 0% | 3,180 | 4,899 | +54% | 0 | 0 | — |
case-12 | fail→fail | 15,168 | 7,982 | -47% | 1 | 1 | 0% | 2,499 | 4,535 | +81% | 0 | 0 | — |
case-14 | pass→pass | 15,081 | 8,707 | -42% | 1 | 1 | 0% | 2,326 | 4,346 | +87% | 0 | 0 | — |
case-15 | fail→fail | 10,139 | 9,136 | -10% | 1 | 1 | 0% | 1,646 | 4,383 | +166% | 0 | 0 | — |
case-16 | fail→pass | 18,282 | 14,119 | -23% | 1 | 1 | 0% | 2,735 | 4,814 | +76% | 0 | 0 | — |
case-17 | pass→pass | 21,631 | 15,631 | -28% | 1 | 1 | 0% | 3,162 | 5,372 | +70% | 0 | 0 | — |
case-18 | pass→pass | 4,152 | 13,140 | +216% | 1 | 1 | 0% | 632 | 3,864 | +511% | 0 | 0 | — |
case-19 | pass→pass | 16,524 | 13,741 | -17% | 1 | 1 | 0% | 2,394 | 5,241 | +119% | 0 | 0 | — |
case-20 | pass→pass | 16,341 | 17,985 | +10% | 1 | 1 | 0% | 2,732 | 6,133 | +124% | 0 | 0 | — |
case-21 | pass→pass | 18,166 | 22,659 | +25% | 1 | 1 | 0% | 3,265 | 6,912 | +112% | 0 | 0 | — |
case-22 | pass→pass | 11,638 | 14,383 | +24% | 1 | 1 | 0% | 1,885 | 5,307 | +182% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 20 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +45 percentage points is the difference between those two pass rates over the 20 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.