Imagine you're a locksmith's apprentice. Your master gives you a ring of old keys and says: "The new lock uses the same pins as the old one — find me a key that still fits." That's the core mechanism of ns2jb. The Switch 2's security is new, but its backward-compatible game library means old code with old bugs still runs inside the new fortress. The exploit chains three separate vulnerabilities across two generations of hardware to go from a BASIC interpreter to native code execution on the Switch 2. The first link is a 32-bit integer overflow in SmileBASIC, a Switch 1 game that embeds a full BASIC interpreter. The REMOVE command checks position + count <= array.length, but the sum can wrap negative at 32 bits, letting the attacker pull 8 bytes of heap metadata — specifically, a self-referential pointer — into the array. That pointer, interpreted as a string length field, gives effectively unbounded read/write within the process. This is a known bug class; smilehax on the 3DS exploited the same codebase years ago. The researcher (with Codex doing bug-hunting) found a new instance of the same structural weakness. The second link is a textbook heap buffer overflow in the Switch 2's built-in filesystem sysmodule (fs). Path normalization for paths containing backslash parent-directory components allocates a 0x301-byte heap buffer but copies the full input without bounds checking. By overflowing into an adjacent BufferedStorage cache entry and setting its dirty bit, the attacker gains arbitrary read/write within the fs process — a privileged, built-in sysmodule. From there, the attacker spoofs SmileBASIC's process ID to 1, granting FullPermission and dumping all ordinary sysmodule code. The third and most elegant link is a logic bug in the loader's signature verification. When loading a program, ldr asks fs to serve the ExeFS filesystem from an NCA archive. fs provides the NCA header hash and developer signature; ldr checks that Nintendo's signature covers the process permissions and developer key, and that the developer's signature covers the NCA header hash. The chain of trust runs: NPDM signature → developer key → NCA header hash → filesystem hashes → executable content. But the filesystem hashes are enforced by fs, not ldr — and the attacker already has arbitrary r/w in fs. They swap the legitimate ExeFS with a temporary filesystem serving arbitrary code while keeping the valid NPDM for the signature check. The second check passes trivially because the developer's signature is verified against the NCA header hash, not the actual served content. The result is a tethered, software-only jailbreak: arbitrary native code execution as any ordinary sysmodule and as the fs built-in sysmodule. The tethered requirement (JetKVMs for USB input, HDMI capture, a Bluetooth switch to press the power button on crash) means this isn't a one-click consumer tool. It's a research harness. Codex — OpenAI's coding agent — was used for bug discovery, offline binary analysis, and even vibecoding a Mario Kart World mod in a day, but the author is explicit that the jailbreak itself took ~2 weeks of directed human effort with Codex as an accelerant. The security implications are significant. Nintendo's Switch 2 deployed PAC, CFI, and XOM — modern mitigations that make traditional ROP and gadget-based attacks far harder. The jailbreak sidesteps all three by avoiding WebKit (which has all mitigations active) and instead chaining legacy game code (no PAC on Switch 1 binaries) with a logic bug that never requires executing injected code in a protected context. The lesson: backward compatibility is a structural attack surface. Every old binary you run inside your new security perimeter is a potential key. The WebKit angle is worth noting for what it reveals about the ecosystem. The author acknowledges that WebKit on Switch 2, hardened with XOM, PAC, and CFI, resisted Codex's automated attempts at exploitation. The Nintendo Bug Bounty Program has likely attracted substantial prior scrutiny of the browser surface. The successful path went around the hard target, not through it — classic adversarial behavior. This is the first publicly disclosed software-only jailbreak for the Switch 2. It demonstrates that AI-assisted vulnerability research can compress timelines substantially — Codex found the SmileBASIC integer overflow, the fs heap overflow victim allocation, and performed offline sysmodule analysis — but the exploit architecture, the insight about backward-compatibility as attack surface, and the loader logic bug required human-directed strategy. The human-AI collaboration model here is more interesting than either component alone.
