Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Classify the dominant artifact, primitive, or objective.
Load the closest offensive-techniques methodology before selecting tools.
Load the closest top-level reference for the dominant primitive (overflow.md, rop.md, heap.md, heap-fsop.md, relro-aslr-relocations.md, sandbox.md, kernel.md, exotic-arch.md, advanced-primitives.md, weird-machines.md, windows-pwn.md).
Load only top-level references in references/.
Choose the smallest tool chain that can produce a validation signal.
Record the exact proof path and stop once the objective is reproducible.
Technique integration
Primary methodology to load:
reversing-technique
vuln-exploit-technique
fuzzing-technique
Use these as decision engines. This skill adds challenge-oriented triage and time-boxing.
Tool routing
Prefer these tool families when the corresponding signal appears:
pwntools
gdb
radare2
ghidra
coding/asm-patterns
coding/asm-testing
offensive-coding/rop-development-dev
offensive-coding/heap-exploitation-dev
offensive-coding/shellcode-dev
Tool syntax belongs in the tool skills. This skill decides when a tool family fits and what output should validate progress.
Category-specific quick pivots
Triage binary and mitigations first; prove exact controlled primitive before payload engineering.
Build exploit in stages: local repro, info leak, base calculation, control-flow/data-only effect, remote adaptation.
Keep offsets, libc/loader assumptions, and environment drift explicit.
For stateful exploits, keep leaks, allocator shaping, staging, and the trigger on one connection unless process boundaries are proven irrelevant. Load references/stateful-exploit-campaigns.md when later stages depend on earlier menu actions, sort order, allocator history, or live stack state.
Treat emulators, permutation models, and offline address calculations as candidate generators. Promote a candidate only after a live breakpoint snapshot proves the exact call site, register arguments, target memory, return state, and next reachable stage.
If _dl_fixup, link_map, dynamic tags, forged symbols, or an unresolved PLT entry become an endgame, load references/dynamic-linker-resolver-pivots.md. Prove a reachable lazy call in the final process state before building resolver metadata.
Harness reliability: anchor each response parser on a stable token (the menu reprint), not the input prompt the previous step already consumed — a swallowed delimiter desyncs the next recvuntil and reads as a flaky exploit, usually papered over with long timeouts instead of fixed. Parse binary leaks over raw pipes or the live socket, never a pty (a pty mangles non-printable bytes).
For local ground truth during development — PIE/heap/libc bases and arbitrary R/W to validate offsets before wiring the leak — read /proc/<pid>/{maps,mem} directly instead of trusting the exploit's own parsed values.
If RELRO, GOT/PLT, relocation tables, or leakless partial overwrites appear, load references/relro-aslr-relocations.md before deciding the final target.
If shellcode has a byte blacklist (filter function rejecting specific bytes), load references/shellcode-filtering.md first — decode the blacklist semantics, find safe XOR/ADD encoding, use register-based string construction to avoid blocked opcodes and string literals.
If the binary stacks Clang -fsanitize=safe-stack, cross-DSO CFI, a hand-rolled software shadow stack, and/or a canary on the unsafe stack (i.e. an overflow that can't reach the return address), load references/safestack-cfi-shadowstack.md. Each mitigation is defeated leak-free by attacking the runtime data they share (the TLS block, the callee's saved-register window, the jump-table-base register) rather than fighting it head-on: unsafe-pointer relocation, canary self-reference/overwrite, CFI base-register rewrite, shadow-stack bare-gadget pivot, sustained main-re-entry loop. Critically, the unsafe-stack neighbourhood depends on RLIMIT_STACK: ulimit -s unlimited (typical socat deploys) puts the writable TCB directly above the unsafe stack so the overflow reaches __safestack_unsafe_stack_ptr/canary — a WSL or default-stack /proc/pid/maps showing libc abutting the unsafe stack is a top-down-mmap artifact, NOT the remote. Reproduce with patchelf+ulimit -s unlimited, or length-sweep the live service.
Quality gates
No claim without a validation signal: recovered secret, replayed exploit, decoded artifact, reproduced model behavior, or corroborated evidence.
Do not brute force before representation, constraints, and success oracle are known.
Keep a pivot ledger: hypothesis, evidence, result, next shortest path.
Preserve coverage by using the top-level references in references/ and keeping cross-links between them consistent.
Keep challenge/platform/competition names out of notes and generated reports.
Failure modes to avoid
Never declare a vector "impossible" or "dead" by reasoning. Impossibility and reachability are decided by dynamic evidence only — code-path enumeration is always incomplete, so deductive impossibility proofs are almost always wrong. A "dead vector" claim requires a failing live test, not an argument. If you write "X can't work leaklessly/deterministically" or "nothing reaches Y", rewrite it as "X untested — run the discriminating experiment," then run it. Repeated impossibility arguments are the signal to hook everything and fuzz, not to write a longer proof.
Prove reachability by hook + fuzz, not deduction. For "does anything call Y", "can malloc return Z", "is this slot/gadget reached": breakpoint Y (and every candidate call site at once) and vary inputs — huge/boundary numeric tokens, error paths, allocation/scratch-buffer thresholds, locale. Internal parser and buffer growth introduce calls absent from normal input. Only "no hit after adversarial fuzzing" counts as unreachable.
Drive to end-to-end execution early, then fix the latest crash. After components verify in isolation, fire the whole chain and iterate: at each fault read x/i $pc, rsp & 0xf, and the faulting field, fix that one point, re-run. Late bugs — resolver-vs-call stack parity (movaps), controlled bytes overlapping a lock/pointer/terminator field, parser overwriting your string — surface only in a full run, never in isolated tests.
Keep a facts ledger and re-apply every proven fact to each later sub-chain. Carry forward every runtime fact (live offsets, register snapshots, alignment quirks, call-site parity, deterministic pointers). A fix found in one context is usually needed again in the final chain; forgetting your own earlier finding is a self-inflicted loss.
Re-read this skill's own pivots when the endgame matches one. If loader/_dl_fixup/forged-symbol, stateful, or partial-overwrite signals appear, load the matching reference before concluding the path fails — the guidance to prove reachability lives there, not in your head.
Reproduce the REMOTE, not your dev box, before trusting a boundary or a "dead" verdict. Kernel, RLIMIT_STACK, and libc differences move mmap adjacencies and mitigation reachability — WSL's top-down (default-stack) layout vs the remote's ulimit -s unlimited bottom-up layout is a classic false wall that makes a solvable overflow look "bounded/impossible". patchelf a copy onto the target libc and run under the target's stack limit, or length-sweep the live service, before any impossibility claim.
An INPUT-ONLY reproduction is the only proof a primitive works. Never accept a helper's, a workflow agent's, or a gdb set/hand-painted sim's "primitive works / doesn't" verdict — reproduce it delivered purely by attacker input. A local PTY in cooked mode silently mangles 0x7f/0x0a/0x11/0x13 bytes in pointer payloads, which fakes an "exploit never fires" negative (drive local dev over a pipe, not a PTY); a gdb set variable that hand-writes a shadow slot or register is not a delivered primitive.
Resources
references/advanced-primitives.md — advanced edge cases: seccomp oddities, VM/JIT/interpreter bugs, runtime pivots, and data-reinterpretation tricks that do not fit cleanly under the core references.
references/dynamic-linker-resolver-pivots.md — load when a leakless chain redirects loader metadata, forges dynamic symbols, triggers lazy resolution through a library path, or chains multiple resolver calls.
references/exotic-arch.md — RISC-V, ARM32, ARM64, and MIPS exploitation notes: register conventions, shellcode, gadget patterns, and architecture-specific pitfalls.
references/rop.md — x86-64 ROP and shellcode flow: leaks, pivots, chain assembly, and constrained environments.
references/safestack-cfi-shadowstack.md — Clang SafeStack + cross-DSO CFI + software shadow stack + canary as a composed mitigation: leak-free breaks (unsafe-stack-pointer relocation, canary self-reference/overwrite, CFI jump-table-base rewrite, shadow-stack bare-gadget pivot, sustained re-entry loop with 16-byte-alignment + locale-forge fixes) and the RLIMIT_STACK/WSL layout trap.
references/sandbox.md — restricted-environment escapes, proc-based pivots, and command-execution constraints.
references/shellcode-filtering.md — byte-blacklist bypass: blacklist semantic analysis, XOR/ADD register-based encoding, alternative syscall construction, blocked instruction substitutions, and ORW shellcode template for filtered execve.
references/stateful-exploit-campaigns.md — load when an exploit depends on same-process state, allocator/call-site drift, deterministic data permutations, staged pivots, or probabilistic remote retries.
