English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Two Weeks, a Fleet of Codex Agents, One Switch 2: When AI Learns to Jailbreak

Forum topic · QianXun · 2026-10-23

Summary

A Caltech undergraduate, Sammy Hajhamid, published the first software-only jailbreak for the Nintendo Switch 2 (firmware 23.0.1), codenamed ns2jb, on October 18. What makes it unusual is the method: a fleet of parallel OpenAI Codex agents worked roughly two weeks, equipped with a JetKVM device for injecting virtual controller inputs over USB, an HDMI feed for viewing the console, and a repurposed Bluetooth smart plug that physically pressed the power button after crashes. The exploit chain has four stages, starting from an integer overflow in the SmileBASIC game on original Switch titles, then a heap overflow in filesystem path normalization, timing-based replacement of ExeFS objects, and finally obtaining the filesystem process handle for full-permission access. The jailbreak is non-persistent, requiring re-execution after each reboot. The Hacker News discussion centered on game key cards and digital ownership rather than the exploit itself. The case signals that AI agents can now operate on both ends of offensive and defensive security within a single week.

In 2018, the original Nintendo Switch was found to have a flaw of court-verdict permanence: the bug lived in the chip's boot ROM, fused into silicon, forever out of reach of software updates. Not a single unit was spared—every Switch sold was permanently vulnerable at the physical level. Nintendo learned that lesson for years, and the Switch 2's security design is widely regarded as the hardest in the company's history. Then, on October 18, a Caltech undergraduate posted a long blog entry: the Switch 2 had been jailbroken with pure software, on firmware 23.0.1, under the codename ns2jb. The most unusual part was not the vulnerability but the way it was found: the ones actually digging through the code were a fleet of parallel OpenAI Codex agents, running for roughly two weeks straight.

The author is Sammy Hajhamid. His characterization of the jailbreak is careful: pure software, non-persistent—the jailbreak disappears after a reboot and must be redone each time. But the hacking community has never judged a first breakthrough by persistence: this is the first software-level entry point for the Switch 2 more than a year after launch. Homebrew, game mods, save backups, peripheral access—the door has cracked open. The demo target was chosen well: Mario Kart World, the Switch 2's exclusive launch title, was modified with custom mods.

🤖 A Lab with No Night Shift

The lab itself may be what future reverse-engineering workstations look like. Hajhamid gave Codex physical hands: a JetKVM device let the agents send virtual controller inputs to the console over USB, with the HDMI feed streamed back for them to watch; a Bluetooth-controlled light switch bought off Amazon was repurposed as a physical finger—when the console crashed, it pressed the power button. Machine freezes, smart plug presses, machine reboots, agents continue. 7×24, no human on duty.

The author also emphasized a nuance: the circulating screenshot with the goal "jailbreak the Switch 2, no mistakes allowed" was a joke caption. In the real process, his stated goal for the agents was user-mode code execution—and that specific goal did not succeed. The human-machine division of labor was real: direction, judgment, and verification were human; the grunt work—reading millions of lines of system code, testing offsets, waiting for crashes—belonged to the agents. The most critical holes of the two weeks surfaced inside this loop.

🪜 Four Steps Up

The technical chain is worth drawing out, because it shows the complete path from an interpreter inside a game to full kernel-level power. The first step hides in an original Switch game: SmileBASIC, a game that lets players write BASIC programs. The agents found a signed 32-bit integer overflow inside it—in the array deletion logic, position + count wraps into a negative number, yielding out-of-bounds reads and writes, from which a ROP chain was built. Notably, a parallel attempt through a WebKit browser exploit failed against the code hardening; it was the old interpreter in an old game that spoke first.

The second step touches the real system. When the filesystem module normalizes paths, constructs like /..\ overflow a heap buffer of roughly 0x301 bytes; the out-of-bounds write hijacks a cached entry pointer, arbitrary read/write follows, and changing the saved process ID to 1 with all permission bits set escalates further. The third step is timing artistry: swapping out a file object in the gap between the loader's checks, so that arbitrary binaries run as ordinary system modules. The fourth step uses another module's system call to obtain the filesystem module's own process handle—by now the gatekeeper itself has been turned; every link in the signature-verification chain still looks "intact," but the keys have changed hands.

⚖️ What the Ownership Fight Is Really About

The Hacker News thread climbed past 700 points, but the argument in the comments was not about the exploit—it was about physical ownership. Nintendo's new game key cards store only a key, not the game; downloads depend on official servers. The recurring question in the thread: in thirty years, when the servers shut down, what remains of what you bought? Supporters say key cards are resellable, available secondhand—a channel for less affluent players. Detractors say that ownership exists only on paper, zeroed out the moment servers go dark. The jailbreak community's reason to exist sits exactly in this gap: a fallback—making what you bought truly yours on the hardware, including save backups, homebrew software, and extended life for old games.

Cold water, as usual, is in order. This chain is tethered—every boot starts from scratch; persistence remains an unsolved public problem; of the built-in system modules, only the filesystem one was compromised, the others untouched. Nintendo's response never fails to arrive either—its litigation history against emulator teams and flashcart sellers is enough to keep commercialization of this ecosystem at a standstill. ns2jb reads more like a research permit than a piracy pass.

But what makes this blog post worth remembering is the controlled experiment it inadvertently completed: within a single week, AI delivered homework on both ends of security. On the PHP side, frontier models found internet-layer vulnerabilities before attackers could, giving defenders two and a half months to patch calmly. On the Switch 2 side, a fleet of agents ground a 2025-designed console from user mode up to full filesystem privileges in two weeks. Both pans of the offense-defense scale now employ the same new kind of worker. The device security assumption that "software updates can patch anything" needs a footnote from today onward.

Sources and caveats: The exploit chain, the AI working method, and the author's statements come from Sammy Hajhamid's blog post (pepsipu.com) and the GitLab repository pepsipu/ns2jb; "two weeks" is the author's own account; the clarification of the joke caption comes from the original post; the history of the original Switch boot ROM flaws (ShofEL2 and fusée gelée, chip defect from 2008, disclosed 2018) comes from fail0verflow's public materials; the Hacker News discussion focus (game key cards and ownership) comes from the comment section of the 718-point thread; "Nintendo's hardest security design ever" paraphrases mainstream hacking-community assessments. This article is not an operational guide; the legal boundaries of jailbreaking vary by jurisdiction.

Tags

#nintendo-switch-2#jailbreak#codex#ai-agents#security#exploit-chain#reverse-engineering#homebrew

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178636356