LIVE TICKER
[Live wire]Connecting to the newsroom…
TechnologyPublished 10 November 2026 • 10 min read • 2,025 words • 5 cited sources

Connected-car hack: ForgeRock flaws let researchers unlock and start vehicles

Researchers used hard-coded ForgeRock credentials and web firewall bypasses to reach two automakers' connected-car backends, unlocking doors and starting engines. Both companies fixed the flaws within about two hours.

Connected-car hack: ForgeRock flaws let researchers unlock and start vehicles
Getty Images via Tom's Hardware — view at source ↗

Key facts at a glance

Affected companies
Two automakers, referred to only as Automaker #1 and Automaker #2Sam Curry
Time to fix
About two hours after the report, with credentials rotatedSam Curry
Default console login used
admin:admin on the hidden Apache Felix console inside OpenIDMSam Curry
Regions demonstrated
Australia, the United States and CanadaShubham Shah
Bug bounty paid
None, according to the researchers' interview with iTnewsiTnews
Earlier car brands reached
Honda, Acura, Nissan, Infiniti, Hyundai, Genesis, Ford, Porsche, Kia and SubaruSam Curry

What did the researchers actually do?

Two security researchers spent part of 2026 probing the account systems that control connected cars from two automakers. They published their findings on 7 November. Their conclusion is blunt: a chain of ordinary web flaws let them act as any customer, including sending unlock and engine-start commands.

Sam Curry and Shubham Shah did not attack the cars directly. They attacked the login and identity servers behind the mobile apps. From there, they minted the same tokens an owner's phone receives. Curry describes gaining "remote code execution on the production ForgeRock servers" that handled customer logins.

Remote code execution sounds dramatic, and here it was literal. Curry's writeup shows the pair running commands inside the identity server. They then used that foothold to call local admin APIs that the public firewall blocked. No vehicle network or key fob was involved.

The prize was the full command set. Inside an account, the researchers could see location, mileage, fuel or battery range, tire pressure, door and window state, vehicle health and subscription status. On supported models, they could lock or unlock doors, start or stop the engine, sound the horn, flash the hazards or locate the car.

Curry says the pair demonstrated the chain with a live US account and vehicle, with the owner standing beside it. Shah separately recorded the Australian version on video. The commands worked: the car unlocked, the engine started, the engine stopped and the doors locked again.

Both automakers fixed the problems within about two hours of receiving the report, according to Curry's timeline. They also rotated the exposed credentials and removed the researchers' proof files. That response time is unusually fast, and it is one reason the writeups are worth reading closely.

Which automakers were affected, and why won't they say?

Curry calls the companies Automaker #1 and Automaker #2 and says both asked not to be named. He agreed to that request. His stated reason was practical: he did not want security teams pulled away from other incidents to manage a public naming fight.

iTnews describes the affected company as a carmaker that sells two vehicle brands in Australia. The Register asked Toyota whether speculation about its involvement was accurate. A spokesperson neither confirmed nor denied it. That exchange is not confirmation, and it should not be read as one.

Online observers guessed Toyota after studying a blurred car and an unlock sound in Shah's video. Tom's Hardware noted the footage appeared to show a RAV4. None of that is proof, and the researchers have not budged from their anonymity agreement.

The safe reading is narrower. Two unnamed automakers, one shared identity vendor, and no public attribution. Treat brand guesses as guesses, including the most confident ones in forum threads.

What is not in doubt is the reach of the affected systems. Curry says they served customers in multiple regions. Shah writes that the US, Canadian and European identity environments had received much more hardening than the Australian one.

What is ForgeRock, and why does one login server control so much?

ForgeRock sells identity software that companies use to run customer logins. In both apps, the vendor's platform handled the entire sign-in journey and stored the customer records behind it.

Curry's writeup splits the platform in two. Access Management, or AM, runs the login steps and issues the session. Identity Management, usually called OpenIDM, holds the customer records that AM consults during that journey.

That split explains the blast radius. A record changed in OpenIDM alters what AM will accept at the next login. A token minted by AM is then accepted by the telematics APIs that issue commands to cars.

The two apps looked nearly identical at this layer. Both used the same consumer realm, the same OAuth client name, the same callback address and the same server-driven callback model. Curry describes them as brand-specific entrances to one identity platform.

OpenIDM also bundles a Java runtime with an administrative console called Apache Felix. The console can edit live service configurations. ForgeRock's own hardening guide recommends removing or protecting it, because it authenticates separately from the identity service.

That separation is the crux. The firewall was tuned to protect the OpenIDM endpoints. The console sat on a different path and kept its own login. Bypass the filter, and the console becomes a route into the same process.

For buyers, the lesson is that identity software is infrastructure. It is rarely the product a carmaker advertises, but it decides who can start a car.

How did the Australian attack start with a number plate?

Shah's writeup is the more approachable of the two. It starts with a question most owners never ask: what does the app talk to when you tap unlock?

He decompiled the Android app and traced its calls to a telematics domain. Scanning subdomains turned up an admin host with a plain login page titled TEST HARNESS. That page shipped its entire JavaScript bundle to anyone who loaded it.

The bundle contained API keys. Many admin endpoints accepted those keys with no other authentication. Customer lookups by email worked. So did queries that turned a subscriber ID into a full account record.

The public route in was a number plate. Australian state registration sites return a VIN. An exposed API converted that VIN into a subscriber GUID, and another endpoint returned the account email. Shah summarizes the chain as plate to VIN to GUID to personal data.

The admin tool alone could not drive a car. It could add a secondary user, provision features and generate remote access codes. For commands, the pair needed the identity layer underneath.

That layer was ForgeRock OpenIDM, which stored every customer identity for the platform. A vendor JavaScript file also carried administrator credentials meant for a staging environment. Those credentials worked in production.

Nginx rules blocked the obvious administrative paths. Shah and Curry spent two weeks studying the identity system. They found a path traversal that slipped past the firewall, and an OAuth password grant that let them skip multi-factor authentication. With a reset password and a minted app token, the command API accepted requests.

iTnews reports that one firewall-exposed endpoint returned password and PIN hashes, push tokens, phone numbers and full names for queried accounts. That is the kind of data that turns a single lookup into a broader intrusion.

Why was the American version harder to break?

The North American environment had more filtering in front of it. Shah told iTnews that reaching the same level of access there was "definitely a lot harder" than in Australia.

Curry's writeup walks through the US chain step by step. Blocked administrative paths returned 403 errors, and repeated attempts got the researchers' IP address denied. A crafted path built from a static prefix, encoded slashes and a semicolon got through. The edge saw a static file, while the Java server decoded a management console.

That console was Apache Felix, the component-management console inside OpenIDM. It still accepted admin:admin, a documented default. Curry notes that ForgeRock's own hardening guidance recommends removing or protecting the console.

The console's configuration manager could schedule a Groovy script. The script installed a small servlet, which gave the researchers a stable command channel. From inside the server, they reached local admin APIs that the public firewall kept blocking.

With that access, they changed a customer's email address to an inbox they controlled. They set a new password and captured the six-digit verification code. Then they completed the normal login, exchanged the session for OAuth tokens and loaded the account's garage.

The inbox was purpose-built. The pair registered a small mail service, generated a fresh address for each run, and polled it for the newest code. Their script ignored messages that arrived before the request, so stale codes could not break the flow. Some profiles asked for a second code, and the cutoff moved forward after each accepted message.

The token exchange is worth spelling out. The access token authenticated requests to the mobile APIs, while the identity token named the consumer and supplied the GUID used to fetch the account. Neither value needed anything from the car itself.

Two details stand out. The app used a PKCE challenge value of "plain", which offers little protection if an authorization code leaks. Push notification registration was also client-controlled, so the researchers could register their own token and keep alerts off the owner's phone.

During the test, the owner's app stayed logged in. The researchers restored the email address afterwards. The original password could not be restored because they never knew it, so the owner chose a new one.

The chain also exposed some internal scale. The console reported 361 bundles in total, with 343 active, running on Apache Felix 7.0.1. That level of detail is a reminder that these identity servers are large Java estates, not single applications.

What does this say about connected-car security?

The individual flaws are mundane, and that is the point. The list includes a default console password, credentials in a shipped bundle, a firewall that parsed URLs differently from the server behind it, and an OAuth flow that skipped multi-factor checks.

None of that required a lab, a soldering iron or a compromised dealership. It required patience and an internet connection. Shah's work began with a public number plate lookup that anyone could run.

The identity layer is the soft target. Modern cars are controlled through cloud APIs, and those APIs trust tokens. Whoever can mint a token for a customer can act as that customer, both in the app and at the vehicle.

Filtering is not a substitute for hardening. Both companies had invested in web application firewalls, yet the services underneath stayed reachable through encoding tricks. Curry's writeup shows the edge and the application disagreeing about what a URL meant.

Neither layer was obviously broken on its own. The firewall matched dangerous patterns, while the server normalized URLs before routing them. The gap between those behaviors was enough. Curry documents several variants that failed before one satisfied both parsers.

This is not the first round of car research for either man. Curry lists Honda, Acura, Nissan, Infiniti, Hyundai, Genesis, Ford, Porsche, Kia and Subaru as brands he and others have reached in recent years. The pattern is industry-wide, not confined to one brand.

Disclosure norms also lag. iTnews reports that the pair received no bounty or public recognition. A two-hour fix is commendable, but it also suggests the problems were easy to spot once someone pointed at them.

The gap between fix and publication is normal. Curry's timeline puts the testing and reporting in early 2026, with the writeups appearing on 7 November. That interval gave both companies time to patch and rotate credentials before the technique became public.

The lesson travels to any organization running the same identity platform. An exposed management console or a stray credential can undo a great deal of perimeter work.

  • •A shipped JavaScript bundle contained staging administrator credentials that worked in production.
  • •A firewall and a Java server disagreed about the meaning of the same encoded URL.
  • •A management console accepted its documented default password.
  • •An OAuth password grant allowed multi-factor authentication to be skipped.
  • •A client-controlled push registration let an attacker suppress the owner's alerts.

What should owners and buyers do now?

Current owners of the affected brands have little to do. The flaws were fixed and credentials were rotated. Neither researcher reports evidence that anyone else used the chain, and neither writeup calls for owner action or a recall.

Buyers should ask harder questions before choosing a connected car. Does the automaker run a vulnerability disclosure program? Does it pay bounties? How fast does it patch identity infrastructure? The answers separate companies that treat security as a process from those that treat it as a launch checkbox.

Owners can reduce exposure in small ways. Use a unique password and app-based multi-factor authentication where offered. Watch for unexpected password-reset emails or sign-in alerts. Keep the app updated, because client fixes arrive through app stores.

The industry lesson is unglamorous. Inventory management consoles, remove default passwords, strip secrets from shipped code, and test how your firewall and your application parse the same URL. Most of this work is configuration, not invention.

Watch three things next. Whether the identity vendor publishes guidance on the path-confusion issue. Whether either automaker says anything voluntarily. And whether other researchers replicate the chain against different identity stacks, which would show the problem is broader than two fleets.

Books behind the context

Sources used for this dispatch

  1. Sam Curry - Hacking Two Automakers: ForgeRock Path Confusion, RCE, and Remote Vehicle Access
  2. Shubham Shah - unlocking any car from X brand in most of the world
  3. The Register - Researchers pwn automakers' Android apps to remotely unlock cars, honk horns
  4. Tom's Hardware - All net-connected vehicles from two major automakers allowed hackers to remotely unlock doors and start the engine
  5. iTnews - Admin creds left in carmaker app opened doors to vehicles

This article is original analysis and synthesis. Facts, figures and quotations are attributed to the sources above. Images remain the property of their publishers and are credited in the caption.

Questions readers ask

Which car brands were affected by the connected-car hack?

The researchers have not named them, referring only to Automaker #1 and Automaker #2. iTnews describes the company as a carmaker selling two brands in Australia, and Toyota neither confirmed nor denied speculation when asked by The Register.

Did the researchers actually unlock and start real cars?

Yes. Curry says they demonstrated the chain on a live US account and vehicle with the owner present, unlocking it, starting the engine, stopping it and locking it again. Shah recorded the Australian version on video.

Is my car at risk now?

According to Curry's timeline, both automakers fixed the flaws within about two hours of the report and rotated the exposed credentials. Neither researcher reports evidence that anyone else exploited the chain.

Was this a physical or keyless-entry hack?

No. It was a software attack on the cloud identity servers behind the automakers' mobile apps, not a relay attack on key fobs or a hardware attack on the vehicle.

Did the researchers get a bug bounty?

No. iTnews reports that the carmaker paid no bounty and offered no public recognition, although it fixed the vulnerability within hours of the report.

What is path confusion in this attack?

It is a mismatch in how two systems read the same URL. The web firewall treated a crafted path as a harmless static file, while the Java server behind it decoded the path as an administrative console.

Topics:connected carscybersecurityForgeRockvulnerability disclosureautomotive softwareidentity management

More from the research desk