As reported by The Hacker News, researchers at NebuSec have published two distinct remote code execution bypasses against Telegram for Android's hastily-deployed widget sandbox — one exploiting a widget bug the seccomp filter does not cover, and another chaining a filtered widget bug with a kernel vulnerability. The demonstration runs on Telegram for Android 13.2.2 (7303), the very build Telegram pushed after an X account named V12 claimed arbitrary code execution from any Telegram message.
Shield53's read: this is less a story about Telegram's engineering and more a case study in what happens when a sandbox is bolted on under time pressure. The two mitigations — isolating QuickJS in its own process and adding a seccomp filter — are the right instincts. But the execution exposes a pattern defenders should recognize immediately.
Vulnerability Snapshot
- CVE
- No CVE identifier assigned as of this writing. NebuSec's write-up names none and does not confirm coordinated disclosure.
- CVSS
- Not assigned. Based on demonstrated impact (sandbox escape, process-level code execution), Shield53 characterizes severity as High pending confirmation of post-escape reach.
- Affected Product
- Telegram for Android 13.2.2 (build 7303) — the first build to ship the QuickJS process isolation and seccomp filter. Earlier builds lack the widget sandbox entirely and are more exposed.
- Exploitation in the Wild
- No confirmed active exploitation. NebuSec ran its proof on Telegram's test server, not production. Technical details are being withheld.
- Patch / Mitigation
- No vendor advisory has been issued. The seccomp filter in published source is set to audit mode — it blocks exec-family calls but only logs the other ~dozens of calls it names (ptrace, bpf, mount, etc.). ioctl is unfiltered entirely.
Why This Matters Beyond Telegram
Three structural issues in this design are worth internalizing whether you defend a messaging app or any product that embeds a JS engine for extensibility:
- Allowlist-by-default is hard; audit mode is a liability. A seccomp filter that logs-and-lets-through every named syscall except exec is not a sandbox — it is a telemetry feed. The comment in source, 'Audit until the logs from the devices...', signals the filter was never load-bearing. Defenders should assume any audit-mode filter provides zero containment.
- Single-process widget multiplexing breaks isolation between bots. Per NebuSec, one QuickJS process serves every widget from every bot. A use-after-free or out-of-bounds write in one bot's widget reads and writes the memory of every other bot's widget in the same process. Process isolation from the host app is real; isolation between widgets is not.
- Engine-level rules are not kernel-level rules. Telegram's widget documentation promises 'no network access' and a sandbox, but those guarantees live in QuickJS code. A single memory-corruption bug in QuickJS erases them. This is the classic gap between documented policy and enforced policy — the same gap that has burned every plugin host from browser engines to Electron apps.
The demonstrated bypass opens the phone's calculator — a benign target chosen for proof. The same primitive that calls am start can call anything ioctl or the unfiltered syscall surface exposes. Defenders should not calibrate to the calculator; they should calibrate to the syscall surface.
Who Is at Risk
Any Telegram for Android user who interacts with bot-sent widgets is potentially exposed. The threat model widens significantly for organizations that use Telegram bots for operational coordination, customer support, or content delivery — a single malicious or compromised bot could weaponize widgets against every user who renders them. Users on builds prior to 13.2.2 (7303) have no sandbox at all and should be treated as fully exposed.
Shield53 Recommendations
Immediate Actions
SCMP_ACT_ERRNO mode, not SCMP_ACT_LOG. If MDM supports it, block the Telegram app update channel from enabling widget rendering.am start, /system/bin/sh) and for unexpected ioctl-heavy activity from the QuickJS process. These are strong indicators the sandbox has been escaped.Broader Lesson
The Telegram case is a microcosm of a recurring industry pattern: a feature ships, a researcher demonstrates RCE, the vendor adds a sandbox in days, and the sandbox is bypassed in hours. The lesson for defenders is to never treat a newly-added sandbox as a finished control. Sandboxes earn their reputation over years of adversarial review, not over a release cycle. Until the filter is enforcing, not logging, this remains an open wound.