Security1 distinct publisher2 min readPublished
CERT/CC disclosed an unauthenticated file read and a code execution path in mwEmbed with no fix available. The advisory's only product is a list of network controls you apply yourself.
The Watch · Security desk

Compiled by The WatchSomething wrong?How this is made
The severity figures in circulation belong to the finder. Gerjan Wemekamp of AndDone, credited with both reports, scored CVE-2026-19912 at 10.0 and CVE-2026-19913 at 9.1 and marked them reporter-assigned in his writeup; CERT/CC published no score of its own, and there was no NVD record for either identifier as of 25 August 2026 [14]. Triage tooling that keys off a published severity will therefore return nothing for either bug [16].
The mechanism is one sink reached two ways. `mwEmbedLoader.php` takes the `ServiceUrl` parameter as the target for backend API requests, and `KalturaClientBase` hands whatever comes back to PHP's `unserialize()` without checking the source, the scheme, or the content [7]. Point it at a `file://` path and the server fetches a local file instead of an API response; deserialization fails, and the raw bytes of the file are reflected to the requester inside the error message [8]. Wemekamp used that to read `/opt/kaltura/app/configurations/local.ini`, which holds plaintext database connection strings, admin and console passwords, and internal host references [9].
The second bug reuses the same deserialization and adds a write. `uiconf_id` is appended to the cache folder path without sanitisation, so traversal sequences steer the deserialized payload out of the cache directory and into a web-accessible one, where requesting it directly runs it as the web-server user [10]. Wemekamp notes the file-drop step depends on the file-based cache backend, which is Kaltura's default, and that a memcache-only configuration may suppress that particular path without making the deployment safe [11].
That asymmetry matters for anyone building the compensating controls. Restricting outbound access from the application server is on the advised list precisely because the code execution chain needs an outbound fetch to pull the payload [12]. The file read needs no egress at all, because the fetch target is the local filesystem [18]. Egress policy buys you the worse half of the problem and none of the credential disclosure.
Inventory is the other soft spot. CERT/CC lists html5lib v2.45, v2.103 and earlier, and other v2.x releases that expose the vulnerable endpoint [13]. The trailing clause means a version match will not settle whether a host is affected; the endpoint's presence will. And with six advised actions and zero fixed versions [17], the operator running html5lib is now the de facto maintainer of a component whose upstream did not answer.
One caveat on sourcing: the account here rests on a single published report of the CERT/CC note and the researcher's writeup, with no vendor comment in it [4].
Ranked by verification strength, evidence, and original report placement.
The CERT Coordination Center disclosed two unpatched vulnerabilities in Kaltura's HTML5 video player library that allow a remote, unauthenticated attacker to read arbitrary files from a server and execute code on it.
The flaws, CVE-2026-19913 and CVE-2026-19912, both stem from the same unsafe deserialization in the mwEmbedLoader.php endpoint of the mwEmbed player library, which Kaltura also distributes as html5lib.
No patch is available, and CERT/CC said it was "unable to reach Kaltura to coordinate these vulnerabilities."
CERT/CC said that because the affected endpoint is also exposed on Kaltura's shared, multi-tenant CDN infrastructure, the vulnerabilities affect not only individual customer installations but every tenant served by those shared hosts.
mwEmbedLoader.php accepts the ServiceUrl parameter and uses it as the target URL for backend API requests; the KalturaClientBase PHP client fetches whatever that URL returns and passes it to PHP's unserialize() without validating the source, the scheme, or the content.
Supplying a file:// path causes the server to fetch a local file rather than an API response; the deserialization attempt fails and the raw bytes of the fetched file are reflected back to the requester inside the resulting error message.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Coordinator note plus researcher chain and independent repo check, minus any scored or end-to-end validation on current code
The technical account is unusually specific: a named coordinator advisory with affected versions, a parameter-level data flow from ServiceUrl into unserialize() and from uiconf_id into a cache write, a named reporting researcher with a published writeup, and the publisher's own repository verification that the vulnerable file is byte-identical across 21 release refs spanning 2015 to August 2026. Two gaps hold the score below high confidence: neither CERT/CC nor NVD published a severity assessment, and the researcher states the end-to-end web shell drop was demonstrated on a 2019 docker image rather than the current release.
Wide latent exposure documented, no observed exploitation and no remediation path shipped
Adoption of the risk in the real world is documented only as exposure, not as activity or response. The vulnerable code is present across a decade of releases and CERT/CC says the endpoint is exposed on shared multi-tenant CDN hosts as well as customer installations, but there is no reported exploitation, no KEV listing, no vendor patch or statement, and no count of affected installations or tenants. What exists to be adopted is a set of operator-applied compensating controls, and the sources report no uptake of them.
Reporter-assigned 10.0 and 9.1 run ahead of independent validation
Slightly overstated rather than sensationalized. The severity anchors circulating with this story are self-assigned by the reporting researcher, with no coordinator or NVD figure, and the full exploitation chain was only demonstrated on a 2019 container image. Offsetting that, the article itself is explicit about scope limits, publishes the researcher's memcache caveat, and backs the persistence claim with its own repository check, so the gap is modest.
Reporter self-scoring and vendor silence, offset by a neutral coordinator and disclosed limits
Two identifiable incentive pressures: the reporting researcher, affiliated with a security firm, published both the writeup and the severity scores that frame the story, and the vendor, unreachable by the coordinator, contributed no counterweight. Mitigating factors are that CERT/CC is a non-commercial coordinator, the researcher volunteered explicit scope and configuration caveats, and the publisher added an independent code check rather than relying solely on the writeup.
Detailed and internally checked, but single-publisher and unscored by any independent body
Confidence is held mid-range by sourcing concentration: both cluster items are the same article from one publisher, so there is no cross-outlet corroboration, and no vendor, coordinator score or NVD record is available to test the severity framing. Against that, the underlying primary sources are named, the mechanism is specific enough to be independently checkable, the publisher performed its own repository verification, and the article states clearly what was and was not demonstrated.
build
Calix gateway takes UPnP orders from the internet, and there is no patch to install1 distinct publisher
security
Vulnerability disclosures bent upward in 2026. Algorithm records did not.1 distinct publisher
build
Fabricated SQLite CVEs cleared NVD, CISA ADP and Red Hat before anyone ran the code1 distinct publisher
build
Six MariaDB versions, one real difference: the only reason to leave 10.6 is the July 2026 clock1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
2 articles · August 26, 2026