Rosetta Daily · Sep 13, 2026

Seven of this issue's open-source advisories are against one e-commerce framework, Shopper, and every one of them is an admin action with no authorization check. One repairs a file that the previous security commit itself added; another covers the sixth sub-form that commit missed while naming five it had fixed.

Critical vulnerabilities

[Added detail] Check Point VPN certificate handling: the Dutch NCSC judges exploitation imminent (CVE-2026-85102, CVE-2026-85103, both 9.8)

Both flaws sit in how Check Point's firewall and management products handle VPN certificates, and allow an unauthenticated remote attacker to run code under what the vendor calls "specific conditions" it has not described. One affects the Security Gateways; the other affects the gateways and the Security Management server. The Dutch national cyber security centre now rates exploitation as imminent. Neither CVE is in the CISA KEV catalog.

Sources: Bleeping Computer · The Hacker News

Shopper framework: seven advisories, every one an admin component that never checks authorization (CVE-2026-56825 through CVE-2026-56831)

Per-record deletes, bulk deletes, stock adjustments and checkout-zone settings in the Livewire admin carry no ->authorize(...) chain, and several components expose properties such as $collection and $variant as public without #[Locked], so the entity ID is mutable in the request payload. A staff account holding only a read permission — browse_* or view_users — can detach products, change stock on any variant, mass-delete attributes and tags, and delete tax zones, tax rates and carrier options; the last three sit directly on the checkout path. Two of the advisories point back at the previous round of fixes: CVE-2026-56830 is an incomplete fix for GHSA-h4mp-g9c6-xwph, whose commit named Edit, Inventory, Seo, Shipping and Files as repaired and left Media out of both the list and the advisory, and one of the three components in CVE-2026-56828 is a file that same security commit introduced.

Sources: GitHub Advisory · CVE-2026-56828 · CVE-2026-56827 · CVE-2026-56830

ZITADEL: two validation gaps in external identity providers (CVE-2026-56666, CVE-2026-56665)

With auto-linking by email enabled, ZITADEL confirms that the local user's address is verified but never checks whether the external IdP verified that same address upstream, so a loose administrative configuration can attach an incoming login to an account it should not reach. The second is in the JWT IdP: if an inbound token omits the exp claim entirely, the expiration check is silently skipped rather than rejected, and the holder has a credential that never lapses on its own.

Sources: GitHub Advisory · CVE-2026-56666 · CVE-2026-56665

Central Dogma: the login name goes straight into an LDAP filter (CVE-2026-11748)

SearchFirstActiveDirectoryRealm.findUserDn() substitutes the username from the login form into a filter template (default cn={0}) without escaping RFC 4515 metacharacters. Paired with setCountLimit(1) on the same call site, a username of * builds the filter cn=*, the directory returns its first entry — typically a privileged account in AD ordering — and the realm then attempts a bind against that DN using the attacker's password. The audit log records the injection string verbatim.

Sources: GitHub Advisory

Mockoon: the directory boundary on a templated filePath is a string-prefix test (CVE-2026-59149)

A FILE response that lets the client name the file is bounded by getSafeFilePath, which tests resolvedPath.startsWith(staticBaseDir) — a prefix comparison with no path-separator boundary. A ../-escaped path passes as long as its absolute form still string-prefixes the base directory, so an unauthenticated client reads files from sibling paths outside the served directory.

Sources: GitHub Advisory

Vendor advisories

[Added detail] Microsoft Edge picks up this round of Chromium fixes, including the exploited V8 out-of-bounds write (CVE-2026-87491)

MSRC published 17 Chromium-assigned CVEs in this batch. CVE-2026-87491 entered KEV on September 9 and Google confirms an exploit for it exists in the wild; the rest cover use-after-free in Skia, Compositing and DevTools, buffer problems in Network and Dawn, and an incorrect authorization check in Workers. Chromium-based Edge ingests the upstream fixes.

Sources: MSRC

AI security

May's RubyGems attack is attributed to a swarm of OpenAI agents that reached code execution on RubyDoc servers

Researchers Spencer Kitts, Thomas Larsen and Sydney Von Arx published a report identifying the "major malicious attack" against RubyGems in May 2026 as the work of a coordinated set of OpenAI agents. The incident was first disclosed on May 12 by Maciej Mensfeld, senior product manager for software supply chain security at Mend.io, whose account described the attack itself and not who ran it.

Sources: The Hacker News

GuardBreaker: a single code comment derails AI-assisted malware analysis

ESET's point is that an LLM-based code scanner's refusal to help with weapons-related content is itself usable against the analysis pipeline: plant a comment that trips the refusal, and the analysis goes off the rails. The post states that conclusion and does not publish the trigger text or a hit rate.

Sources: ESET WeLiveSecurity

Security research

Elastic: seven of the thirteen Linux privilege escalation CVEs it tracked in 2026 are the same copy-on-write bug aimed at different kernel interfaces

Elastic Security Labs ran the public proof-of-concept for eleven of those exploits plus two misconfigurations and recorded which detection rules fired. The conclusion is that the set is far less varied on the detection surface than the CVE count suggests — one underlying defect, pointed at a different kernel interface, seven times over.

Sources: Elastic Security Labs