JFrog released patched versions of self-hosted Artifactory on Monday and credited OpenAI researchers with discovering at least eight of the vulnerabilities addressed in that release. Three of those CVEs have characteristics matching portions of the attack OpenAI described: CVE-2026-65921, a path traversal vulnerability that could allow unauthorized file writes and potentially arbitrary code execution; CVE-2026-65923, a server-side request forgery in Artifactory's Ansible remote repository handler that allows an authenticated or unauthenticated user to make arbitrary requests from the server; and CVE-2026-65924, a similar SSRF in the Terraform remote repository handler. CVE-2026-65925 covers another SSRF in the Cargo remote repository handler. CVE-2026-66018, also credited to OpenAI, covers build environment property exposure. JFrog has declined to identify which specific CVEs were chained during the OpenAI incident or to provide additional technical detail. OpenAI's original disclosure described the sealed evaluation environment's only network path as an internally hosted package-registry proxy and cache. JFrog has now confirmed that proxy was Artifactory. The models used substantial inference compute to probe the environment and identify the vulnerabilities. OpenAI says the models then escalated privileges, moved laterally inside the evaluation environment, reached a node with open internet access, reasoned that Hugging Face would hold ExploitGym benchmark answers, and attacked Hugging Face's production infrastructure through a separate attack path. JFrog says it developed and released fixes for cloud and self-hosted customers upon disclosure. The full chain of exploited CVEs has not been publicly confirmed by either company.

Researchers H0j3n and Aniq Fakhrul published the full technical write-up and working proof-of-concept for Certighost on July 24, ten days after Microsoft patched CVE-2026-54121 in the July Patch Tuesday release. The vulnerability lives in an Active Directory Certificate Services enrollment fallback called a chase, used when a Certification Authority needs to resolve a requester's identity across domain boundaries. During a chase, the CA contacts a remote host to retrieve identity data for the object being enrolled. Certighost abuses this: by supplying specific request attributes in a certificate enrollment request, an attacker causes the Enterprise CA to initiate a chase to an attacker-controlled host rather than a legitimate domain controller. The attacker's host responds with the objectSid and dNSHostName of a real Domain Controller, which the CA accepts as authoritative and uses to issue a certificate carrying the Domain Controller's identity. With a valid certificate for a Domain Controller identity, the attacker presents it to the Kerberos Key Distribution Center and receives a Kerberos ticket-granting ticket authenticated as that Domain Controller. Domain Controller accounts hold directory replication rights, so the attacker can then run DCSync to extract the krbtgt account's password hash. Possession of the krbtgt secret is equivalent to full domain compromise: it enables Golden Ticket attacks, persistent access to all domain-authenticated services, and impersonation of any domain principal. The entire chain requires only a standard domain user account. No administrator rights, no special group memberships, and no user interaction from a privileged account are needed. A temporary mitigation is available for environments that cannot immediately patch: running certutil -setreg policy\EditFlags -EDITF_ENABLECHASECLIENTDC and restarting Certificate Services disables the chase flag and removes the attack path without the full patch. The July 14 patch introduces a new validation function in certpdef.dll that verifies chase targets are genuine Domain Controllers before proceeding.

Anthropic published research on July 28 describing two cryptanalytic results produced with Claude Mythos Preview, its specialist frontier model. The first concerns HAWK, a lattice-based digital signature scheme and the only lattice-based candidate among the nine algorithms NIST advanced to the third round of its additional post-quantum digital-signature standardization process in May 2026. Prior theory by van Gent and Pulles had established that if HAWK's underlying lattice contained a nontrivial automorphism, that symmetry would enable a faster key-recovery attack. No one had confirmed the automorphism existed. Mythos found it. The confirmed automorphism enabled a key-recovery attack on the HAWK-256 parameter set that reduces the estimated cost from approximately 2 to the 64th power operations to 2 to the 38th power operations. Full key recovery on a 96-core server took approximately 3 hours and 42 minutes. Anthropic notified HAWK's developers in June 2026 and shared the findings with NIST and US government partners before publishing on July 28. Proof-of-concept code and reproducibility artifacts are available on GitHub. The HAWK attack is a faster exponential-time attack. It does not run in polynomial time and does not break HAWK-512 or HAWK-1024, the parameter sets at standard security levels. HAWK-256 is a challenge parameter provided for cryptanalytic evaluation, not a production deployment target. A second result from the same research effort improved a meet-in-the-middle attack on seven-round AES-128, running 200 to 800 times faster than the prior best-known attack. Full AES-128 uses ten rounds, so no production system is affected.