eMudhra’s Weak Certificate Keys: Exploitation Evidence and Patching Gaps
- Cargo
- Machine Learning
- Shipped
- Shipper
- Maxime Guerreiro
- Readers
- 475
Cloudflare found forged certificates soon after eMudhra’s weak keys became public, while probes showed many older clients still accepted them.

Eight weak eMudhra signing keys turned a certificate failure into a problem that any party could exploit. The keys were so predictable that they could be recovered from public certificates: one was the number 4, and the others were 56, 66, 115, 155, 156, 230 and 237. Cloudflare’s follow-up describes both what it saw on its network and how many clients appeared to still trust certificates made with those keys.
The distinction from earlier certificate authority compromises matters. Incidents involving DigiNotar, TÜRKTRUST and TrustWave gave particular attackers or organizations access to certificate signing capabilities. Here, the affected intermediate private keys could be derived from public information. Anyone who recovered them could create certificates for any domain that an unpatched client would accept, provided that client did not enforce Certificate Transparency.
eMudhra said it would leave the public certificate authority business and attributed the weak keys to a mistake in a key generation script. It had not shared the script at the time of Cloudflare’s account. The handling of a certificate revocation list also raised questions: eMudhra first signed it with a key that may have been generated correctly, then replaced it with a version signed by one of the weak keys. The circumstances around that choice, and the roles of auditors and cross-signers, remained unresolved.
Cloudflare’s response and network evidence
Cloudflare began patching high-impact systems on October 19, 2026. Its customer origin-facing proxies, built on Pingora, handle most of the company’s outbound connections. Those systems were updated within hours to reject the weak eMudhra keys.
In parallel, Cloudflare used AI-assisted source code analysis to identify other systems that trusted the keys. The analysis was completed within four hours. Teams then reviewed preliminary findings, with immediate fixes requested for high-impact systems and lower-impact remediation assigned for the end of the week. At the time of the update, all high-priority systems had been patched, while a small number of confirmed, low-impact tickets remained open.
Cloudflare added logging to Pingora to record outbound TLS handshakes involving certificates signed by the weak keys. Because the service handles most outgoing TLS connections from Cloudflare machines, this gave the company a way to look for attempted interception. Logging was fully deployed on October 21.
During the following week, Cloudflare recorded about 22,000 attempted handshakes with certificates using the weak keys. Roughly two-thirds involved test sites such as emudhra.we-have-a-problem.com. Almost all of the rest went to servers that appeared to have set up their own certificates using a weak intermediate. Cloudflare’s patched systems rejected every handshake that attempted certificate verification.
About 160 handshakes merited closer inspection. Their certificates named domains belonging to other organizations in the Subject Alternative Name field, even though those names did not match the requested Server Name Indication. Some were associated with consumer IoT domains. The certificates had been created after the weakness became public, and some had validity periods longer than two years, beyond the maximum term a public certificate authority may issue.
Those details show the keys were used to create forged certificates. They do not, by themselves, show interception of Cloudflare connections. The certificates looked like default certificates served by the destination servers; they named domains different from the ones requested and would therefore fail hostname checks. In several cases, Cloudflare saw the same certificate from connections in many locations, which points to the destination server supplying it rather than an intermediary on one network route.
One forged certificate was dated October 20 at 18:19 UTC, less than an hour after Mozilla’s bug report became public. Cloudflare had not found evidence that the weak keys were used to intercept its connections or other indirect signs of targeted compromise. That finding cannot rule out carefully selected interception attempts, which may not produce the same visible signals.
What the visitor probes show
Cloudflare cannot directly check whether a visitor trusts the weak keys without disrupting that visitor’s connection. Instead, it used a fraction of its challenge pages to trigger two browser probes: one to ca1.cloudflarecaptcha.com and another to ca2.cloudflarecaptcha.com. Initially, the first endpoint presented an eMudhra-signed certificate and the second a normally signed certificate for comparison. Cloudflare recorded whether each probe succeeded. The experiment collected 50 million probes over the week.
The results need careful interpretation. Challenge page traffic is highly skewed and includes an unknown number of bots; only clients that run JavaScript send the probes. The test setup also changed several times, affecting what the results can be compared against.
- On October 23 at 15:30 UTC, Cloudflare changed the control endpoint to serve an eMudhra certificate without signed certificate timestamps (SCTs). Acceptance differed somewhat, but not substantially.
- At 16:30 UTC that day, an unrelated issue prompted a rollback of the challenge platform. Probes resumed on October 27 at 10:00 UTC.
- On October 27 at 16:37 UTC, Cloudflare added the G3 cross-sign, which some clients did not know, and removed SCTs from both endpoints.
- At 16:49 UTC, ca2 was changed to chain up to the AAA Certificate Services root.
- At 18:18 UTC, Cloudflare raised probe coverage from 1% to 10% of challenge pages on non-paying customer zones.
Among modern browsers, most appeared to have been patched. Older browsers and bots showed more acceptance of the weak chain. Embedded browsers, including Android WebView, lagged: 30–40% still accepted it. Cloudflare saw clear evidence of patching over the first weekend, when the monitoring gap occurred, but little movement afterward.
A change in traffic classification complicates the browser breakdown. On October 29 around 18:00 UTC, Cloudflare deployed detection that identified a large influx of bot traffic claiming to be Chrome 150–154 on macOS. This made the upper chart appear to show a patching shift, but the corresponding increase in acceptance among traffic flagged as bots indicates that the traffic mix changed. Chrome 150–154 on Linux was the exception: acceptance continued to fall, a pattern consistent with further patching, though the changing mix means that conclusion remains tentative.
By network, more than half of probes on large networks still accepted the compromised chain, both for traffic with and without detected bot signals. The trend was flat after some weekend patching. Bots may distort the figures, but they still depend on software that needs updates; their behavior is therefore a useful, if imperfect, warning about patching delays.
Next steps for clients and trust systems
Cloudflare’s guidance is direct: install available updates. Dennis Jackson’s crowdsourced site collects instructions, including steps for macOS developers. Cloudflare customers using products such as WARP or Tunnel can consult its customer guidance.
Cloudflare also points to quick revocation as a way to respond when trust needs to change before ordinary software updates reach every client. It has that capability in some products and wants to extend it across more systems. The upki project reflects the same need: trust stores must be capable of receiving timely updates about revocation and Certificate Transparency keys, rather than remaining static lists of trusted roots. The immediate lesson is practical: clients need a reliable patch path, and trust systems need a way to react quickly when a signing key can no longer be trusted.