Hardware
Tesla: a voltage glitch to break the root of trust
In 2023, four TU Berlin researchers make a Tesla's AMD secure processor stumble by disturbing its power supply through the SVI2 bus. The payoff: arbitrary code on the root of trust, extraction of an identity key, and a jailbreak that software cannot patch.
The root of trust in a Tesla
A recent Tesla's infotainment runs on an AMD Ryzen x86 processor. Its entire startup is locked down by a dedicated component, the AMD secure processor (ASP, also called PSP): at power-on, it verifies the signature of every software stage before launching it. This chain is the system's root of trust. It guards three things: the activation of paid options, the key that proves the car's identity to the Tesla service network, and the encryption of user data.
Break that link and you get everything at once. That is exactly what a TU Berlin team did, in two talks: "Jailbreaking an Electric Vehicle in 2023" at Black Hat USA, then "Back in the Driver's Seat" at the 37th Chaos Communication Congress.
Voltage glitching, explained
A processor assumes a stable power supply. If you make it drop very briefly, by a few tens of nanoseconds, at the precise moment it executes a sensitive instruction, it stumbles: the instruction is skipped or mis-executed. Applied right on the comparison that validates a signature, this fault turns a failure into a success.
The key point: you are not attacking the cryptography. The RSA signature stays sound; you simply stop the processor from verifying it. The comparison "is the computed signature equal to the expected one?" returns "yes" when it should return "no". Unsigned code passes as if it were legitimate.
All the difficulty is in the timing. You have to find the brief window where verification happens, trigger the drop at that instant, and start over until it lands. It is slow to tune; once the window is known, the attack is repeatable.
Where to strike: the SVI2 bus
How do you drop the voltage of a chip you do not open? The trick is that the secure processor commands its own core voltage. It does so through a small serial bus, SVI2, which drives an external voltage regulator (the VRM) on the board. That command bus is accessible, and that is where the attacker taps in.
A microcontroller costing about a hundred dollars, wired to the SVI2, injects a forged command that tells the regulator to produce a very short voltage drop. By aligning that drop with the verification window, you get the fault you want without ever touching the processor's silicon.
This is the heart of "One Glitch to Rule Them All" (Buhren et al., 2021), which the Tesla team reuses: the boot code in ROM loads the next stage from flash memory and verifies its RSA signature; the glitch strikes that verification; a single successful fault is enough to make the ROM run code of the attacker's choosing.
The boot chain and the patched firmware
Once the ROM is past, the remaining stages must be chained together. The team patches the firmware by going back up the chain, from the last stage to the first. The problem: a modified stage no longer has the right hash, and the previous stage, which checks it, would refuse to launch it. The workaround is to combine the two: the glitch neutralizes the ROM's verification, which lets the patched chain run all the way through. The result: a root shell on the infotainment's Linux.
- Remove the MCU board, locate the SVI2 bus and the regulator that powers the secure processor core.
- Wire a microcontroller (about 100 $) to the SVI2, and an oscilloscope to observe startup.
- Pin down the window where the ROM verifies the next stage's signature.
- Inject a voltage-drop command aligned with that comparison, and repeat until the fault lands.
- Get the patched boot stage accepted: the root of trust now runs unsigned code.
- Roll out the patched chain (bootloader, ABL, UEFI, Linux) up to persistent root access.
- Extract the chip-bound secrets: the identity key and the keys that encrypt the data.
What the glitch unlocks
| Capability gained | What it means in practice |
|---|---|
| Persistent jailbreak | Root access on the infotainment's Linux, surviving reboots and updates. |
| Extraction of the identity key | The hardware-bound key that authenticates the car to the Tesla network: enough to migrate, or even clone, a vehicle's identity. |
| Paid options | Heated seats, acceleration boost and other software-locked features, activated without paying (a simple configuration flip, so persistent). |
| User data | Decryption of the personal data encrypted by the system: address book, calendar, location history, service sessions. |
The toolkit
- Microcontroller on SVI2 (about 100 $)
- A Teensy or equivalent, wired to the voltage-command bus. It forges the drop command. This is the heart of the attack, and it costs almost nothing.
- Oscilloscope / logic analyzer
- To observe startup and find the time window where verification happens.
- Soldering iron and physical access
- You have to open the MCU board and tap onto the SVI2 and the power rail.
- Prior art
- The AMD glitch from 'One Glitch to Rule Them All' (2021), then the team's 2025 follow-up, 'Three Glitches to Rule One Car' (ASIACCS), on the driver-assistance computer.
What went wrong
- The root of trust assumes the ROM always enforces its verification. Fault injection breaks that assumption: you do not attack the crypto, you stop the processor from verifying it.
- The processor commands its own voltage through SVI2: that command bus is an accessible injection point on the board.
- The boot code lives in masked ROM, etched in silicon: it cannot be patched by an update.
- The chip does not detect power anomalies (no glitch detector, no double verification): a single fault is enough.
Why it cannot be patched
The boot code lives in masked ROM, that is, etched into the silicon at manufacturing. A software fix cannot add anything to it: there is no way to place a countermeasure where the fault strikes. And since the attack does not break the cryptography but bypasses it, strengthening keys or signatures changes nothing.
The only real fix is in hardware: silicon resistant to fault injection (voltage and clock detectors, double verification, internal voltage ramps). In other words, replacing the chip. Since AMD also supplies other manufacturers, the class of problem goes beyond Tesla. On the server side, Tesla can limit the value of an extracted key by hardening vehicle re-authentication (a zero-trust approach).
Scope and limits
This is not a remote mass-theft: it needs physical access, hardware, and tuning time per unit. But on one count it is worse than theft: the attack breaks a hardware trust anchor and lets you clone a vehicle's identity. And because the flaw is in the silicon, it cannot be fixed by software.
Sources
- Christian Werling (TU Berlin) : page de recherche et publications
- Mayhem Security : Jailbreaking Teslas at Black Hat USA 2023
- VicOne : Tesla jailbreak via firmware patching and voltage glitching
- Hackaday : jailbreaking Tesla infotainment systems
- One Glitch to Rule Them All : fault injection contre l'ASP AMD (arXiv)