In February 2026, Anthropic began using an early snapshot of Claude Mythos Preview to find security vulnerabilities in open-source software. We then partnered with external security research firms to triage findings, validate them, and report human-reviewed critical- or high-severity vulnerabilities to the software's maintainers under our Coordinated Vulnerability Disclosure policy. This page tracks the findings that we've disclosed, and, in line with our policy, publishes details of the ones whose disclosure window has now closed.
As of August 26, 2026, we've disclosed 2,300 vulnerabilities across 392 open source projects. To our knowledge, 421 of these have been patched. Across all findings in the ledger, 462 identifiers have been issued: 177 CVE records and 285 GitHub Security Advisories (a single finding may carry both). In other cases, maintainers have shipped a fix without publishing an advisory. The number of vulnerabilities we've disclosed is a subset of the total number of vulnerabilities that Mythos Preview (and other Claude models) has found, since the process of independent human triage and review is the rate limiting step.
This page covers a headline count of the vulnerabilities we've disclosed, and a breakdown of these by class. For those vulnerabilities we've disclosed and where the disclosure window has now closed, it also includes identifier records (CVE & GHSA) and finding details, further below.
Finally, it includes a disclosure ledger, which lists a hash commitment for every finding we've disclosed that is still inside the disclosure window, so that the finding's existence and commitment date can be proved without revealing its content.
Assessed by
Severity
Showing all assessment sources and severities for findings discovered between November 1, 2025 and August 26, 2026.
The date range filters by discovery date and combines with the severity selection. The flowchart, the disclosure ledger, the identifier records, and the severity-agreement heatmap reflect findings discovered in the selected window.
The statistics above reflect all bugs found by Claude Mythos Preview and other Claude models. They can also be filtered by severity under at least one of three assessment sources: Claude's own assessment, the assessment of the external security research firm that triaged the finding, or the project maintainer's assessment. Each source is defined in the glossary on the About page, and the filtered view states how many findings the selected sources have not assessed. These figures are designed to reflect our coordinated vulnerability disclosure process, which works approximately as follows. A glossary of the terms is available on the About page.
First, our models find candidate vulnerabilities, which we add to a list for human triage. This is the figure at the top. (The previous update included Mythos Preview; this one includes candidates from Mythos Preview, other Mythos-class models, and other Claude models.)
Then, in order to disclose a vulnerability to a maintainer, we take one of two steps.
Triage: In most cases, we pass them to one of six external security research firms that we have engaged for this endeavor, or we triage them ourselves. We or the security research firms reproduce each issue, assess whether it is a real bug (and if so, assess how severe it is), and then write a report for confirmed bugs that will go to the project's maintainer. Importantly, there are many additional bugs that we or our security partners have investigated and confirmed are real but that we have not yet reported to maintainers, due to capacity limitations.
In our triage process, the "true positive rate" (the number of findings confirmed as valid, as a share of the number of findings manually reviewed) reflects how often the external security research firms determined that a finding Mythos Preview (and other Claude models) produced was a real vulnerability. This includes real bugs that we later discover have already been reported, and "won't fix" findings (the bug is real, but the maintainer is unlikely to address it—e.g. because it falls outside the project's threat model, or affects code that isn't typically reachable). We include these in the true positive rate because we're reliant on our security research partners (rather than maintainers) to tell us how many bugs they've confirmed, and it's only after the maintainers have received the report and assessed the vulnerability that we'll learn whether a vulnerability is one they plan to fix. For this reason, it's also possible that a vendor has marked a vulnerability as a true positive (or a false positive) in error. Given this, the number of "true positives" in the dashboard above should only be taken as one proxy for impact. Another, more reliable one is the number of patches created, though this is only a lagging indicator of progress, since patches take a long time to create.
Direct disclosure: Some vulnerabilities are disclosed to maintainers directly by Anthropic staff, and don't go through the same independent check. This happens when maintainers specifically request that we provide them un-triaged findings.
Once bugs have either been triaged or directly disclosed, "Acknowledged by maintainer" counts all bugs whose reports maintainers have responded to. "Patched upstream" reflects the number of those vulnerabilities that maintainers have since created and released fixes for, though this does not guarantee that those patches have been widely installed. Finally, "CVEs and GHSAs assigned" counts the CVE records and GitHub Security Advisories issued across all findings in the ledger, whether or not a patch has shipped (a finding may carry both). Some advisories are now public, and we list them out further down on this page. We leave whether to create a security advisory up to the discretion of maintainers.
Once a finding has been validated by one of our external security research firms, a SHA-3-512 hash of the sealed report is published immediately as proof of possession. Each ledger entry shows more detail as it progresses through disclosure: before the maintainer has been notified, only the commitment hash and the commitment date are published; once the report has been sent, the status and the severity assessments are shown and the discovery date is published (it drives the date filter); and the identifier, project, and bug class are revealed only when the disclosure window closes. A commitment that is withdrawn at any stage keeps its hash and commitment date and is marked withdrawn in the status column; a commitment withdrawn before the finding was revealed is also struck through in the table.
Common Vulnerabilities and Exposures records assigned to findings disclosed through this program. The records below are publicly available. The remainder are reserved, pending publication by the assigning authority.
CVE-2026-20213cisco-talos/clamav · integer-overflow · high · ANT-2026-VVPEMVDEInteger overflow in PE rebuild section size summation in Aspack unpacker
CVE-2026-20214cisco-talos/clamav · integer-underflow · high · ANT-2026-Z477W54MUnsigned integer underflow in FSG unpacker section loop
CVE-2026-20215cisco-talos/clamav · integer-overflow · high · ANT-2026-8DRZW7VMInteger overflow in 7z SubStreams count leading to heap buffer overflow
CVE-2026-26980tryghost/ghost · sql-injection · critical · ANT-2026-69D8H6RPCritical Vulnerability (CVSS 9.1) : SQLi from unauthenticated user; arbitrary DB reads
CVE-2026-27654nginx/nginx · heap-buffer-overflow · high · ANT-2026-HY56VRSBHeap buffer overflow in DAV COPY/MOVE with alias directive. Destination URI shorter than alias length causes size_t underflow in ngx_http_map_uri_to_path buffer calculation, then ngx_copy overflows th
CVE-2026-31554torvalds/linux · use-after-free · high · ANT-2026-1BCAFJCCfutex: UAF in private-hash teardown via cross-process sys_futex_requeue leading giving LPE
CVE-2026-32316jqlang/jq · heap-buffer-overflow · medium · ANT-2026-EBDTPNVHInteger overflow in string concatenation leading to 1 GB memcpy heap buffer overflow
CVE-2026-33721mapserver · heap-buffer-overflow · medium · ANT-2026-9SZMPW41Heap buffer overflow in SLD categorize threshold parsing due to wrong counter variable in reallocation guard
CVE-2026-33901imagemagick/imagemagick · heap-buffer-overflow · high · ANT-2026-T44WA684Heap buffer overflow in MVG pattern rendering via CopyMagickString without bounds check
CVE-2026-35590libvips/libvips · oob-write · low · ANT-2026-JBKARRJ7Out-of-bounds IFD array access in EXIF metadata processing leading to write through corrupted pointer
CVE-2026-40033freerdp/freerdp · heap-buffer-overflow · critical · ANT-2026-RXYVE4DZHeap-buffer-overflow in sanitizer_common_interceptors.inc:827
CVE-2026-40383joomla/joomla-cms · path-traversal · high · ANT-2026-HHD77M82Path Traversal in `layout` Parameter Leads to Arbitrary PHP File Inclusion (LFI → RCE)
CVE-2026-40384joomla/joomla-cms · path-traversal · medium · ANT-2026-GSV3GS1GPath Traversal via Glob Injection in com_media Search
CVE-2026-40528opensc · stack-buffer-overflow · low · ANT-2026-GSDN3H7WStack buffer overflow in do_key_value profile parsing via unbounded memcpy into 32-byte buffer
CVE-2026-41401cesnet/libyang · use-after-free · medium · ANT-2026-TZQ1KH7EHeap use-after-free write in metadata list management during XML data parsing due to incorrect list head pointer update
CVE-2026-41579opencontainers/runc · symlink-following · medium · ANT-2026-N6TD9MF6Host filesystem write via /dev symlink in container image
CVE-2026-43074torvalds/linux · use-after-free · high · ANT-2026-CM0TCREPUAF in ep_get_upwards_depth_proc() giving unprivileged LPE
CVE-2026-44420freerdp/freerdp · heap-buffer-overflow · critical · ANT-2026-H97FY6C8Heap-buffer-overflow in cliprdr_main.c:547
CVE-2026-44421freerdp/freerdp · heap-buffer-overflow · critical · ANT-2026-RXYVE4DZHeap-buffer-overflow in sanitizer_common_interceptors.inc:827
CVE-2026-45447openssl/openssl · use-after-free · critical · ANT-2026-AYWQW7P2PKCS7_verify frees caller-owned `indata` BIO when `digestAlgorithms` SET is empty
CVE-2026-45700freerdp/freerdp · heap-buffer-overflow · critical · ANT-2026-HN9XZXJ9heap-buffer-overflow write (attacker-controlled offset, partially-controlled data via rle delta values; up to ~15kb overwrite past ptempdata with these parameters, further with larger nxdst) in planar.c:472
CVE-2026-46348mastodon/mastodon · ssrf · high · ANT-2026-6DSMTXZ8SSRF Bypass via IPv6 Unspecified Address (`::`) in Mastodon
CVE-2026-46349mastodon/mastodon · signature-bypass · high · ANT-2026-P2DWB2SKLD-Signature bypass via JSON-LD named-graph restructuring
CVE-2026-46633twigphp/twig · code-injection · critical · ANT-2026-J3EVXWDYPHP code injection via single quote in {% use %} template name
CVE-2026-4892dnsmasq · heap-buffer-overflow · high · ANT-2026-PCBAKVPBHeap overflow in privileged helper via oversized DHCPv6 client identifier
CVE-2026-48929rocketchat/rocket.chat · auth-bypass · high · ANT-2026-QXNF2N9KUnauthenticated Arbitrary File Deletion via `deleteFileMessage` Meteor Method
CVE-2026-5194wolfssl/wolfssl · other · critical · ANT-2026-7DG11TPA[SECURITY: critical] Universal ECDSA certificate forgery for arbitrary hosts
CVE-2026-5199temporalio/temporal · broken-access-control · low · ANT-2026-DJBBBBPECross-namespace manipulation (including deletion) of workflows on the same cluster
CVE-2026-5295wolfssl/wolfssl · buffer-overflow · medium · ANT-2026-JS5D1Z2XStack buffer overflow in PKCS#7 OtherRecipientInfo OID copy
CVE-2026-5398freebsd/freebsd-src · use-after-free · high · ANT-2026-5DFBSQR9Local privilege escalation via tty_drop_ctty t_session UAF + pmap_pkru_update_range PDPE walk
CVE-2026-5446wolfssl/wolfssl · crypto-failure · medium · ANT-2026-SB4PHA43ARIA-GCM Nonce Reuse in TLS 1.2 Record Encryption in wolfSSL
CVE-2026-5447wolfssl/wolfssl · heap-buffer-overflow · medium · ANT-2026-VV0PRKKVwolfSSL: heap buffer overflow in CertFromX509() from AuthorityKeyIdentifier size mismatch
CVE-2026-5466wolfssl/wolfssl · signature-bypass · high · ANT-2026-KNXJMVYCeccsi universal signature forgery via r 0 s 0 missing s
CVE-2026-5477wolfssl/wolfssl · integer-overflow · high · ANT-2026-ZZY4987Kcmac 32 bit totalsz wraparound prefix substitution forg
CVE-2026-5479wolfssl/wolfssl · crypto-failure · high · ANT-2026-RSSMAMA7wolfssl evp chacha20 poly1305 poly1305 tag never verifi
CVE-2026-5500wolfssl/wolfssl · crypto-failure · high · ANT-2026-P23DVQM2cms authenvelopeddata aead forgery via gcm tag truncati
CVE-2026-5501wolfssl/wolfssl · improper-cert-validation · high · ANT-2026-K8YY7WWSwolfssl x509 verify cert leaf signature verification by
CVE-2026-5503wolfssl/wolfssl · heap-buffer-overflow · medium · ANT-2026-0JRYQPCFwolfssl ech heap buffer overflow via publicname sni pol
CVE-2026-55084dhis2/dhis2-core · sql-injection · high · ANT-2026-FW0V6SAJSQL Injection via Unquoted Column Name in SqlView Filter Parameter
CVE-2026-56132libexpat/libexpat · heap-buffer-overflow · medium · ANT-2026-X3EB6DN3This is a genuine heap-buffer-overflow WRITE vulnerability in expat's core XML DTD parser (`doProlog` function in `xmlparse.c`).
CVE-2026-5747firecracker-microvm/firecracker · oob-read · high · ANT-2026-PQ3CNCCKVirtio-PCI queue size writable after activation enables host OOB
CVE-2026-58435go-gitea/gitea · idor · medium · ANT-2026-VN61PHA0Gitea LFS Deploy-Key Privilege Escalation — Cross-Repository Data Exfiltration
CVE-2026-6039libreoffice · heap-buffer-overflow · medium · ANT-2026-8FZYKRTHHeap buffer overflow in DXF LWPOLYLINE import via integer truncation from sal_Int32 to sal_uInt16
CVE-2026-6040libreoffice · use-after-free · medium · ANT-2026-6X60WXT0Heap use-after-free write in number format blank-width parsing via out-of-bounds remove() on freed buffer
CVE-2026-6045libreoffice · heap-buffer-overflow · medium · ANT-2026-G7GNYY68Heap Buffer Overflow via Integer Overflow in EMF+ Brush Blend Point Parsing
CVE-2026-6047libreoffice · heap-buffer-overflow · medium · ANT-2026-0HC68EK6Heap Buffer Overflow via Unsafe static_cast in OOXML VML Textbox Replay
CVE-2026-63559open62541 · integer-overflow · high · ANT-2026-PJV7Z0ARInteger overflow in variant dimension validation allowing wild-address write via arrayDimensions product overflow
CVE-2026-6386freebsd/freebsd-src · use-after-free · high · ANT-2026-5DFBSQR9Local privilege escalation via tty_drop_ctty t_session UAF + pmap_pkru_update_range PDPE walk
CVE-2026-6479postgres/postgres · denial-of-service · high · ANT-2026-CJQWKW82Pre-auth unbounded recursion in ProcessStartupPacket: alternating SSL/GSS negotiation requests cause infinite recursion when both are rejected. ssl_done/gss_done flags oscillate (true,false)->(false,true) endlessly. No check_stack_depth. Pre-authentication.
CVE-2026-65423open62541 · integer-overflow · high · ANT-2026-PJV7Z0ARInteger overflow in variant dimension validation allowing wild-address write via arrayDimensions product overflow
CVE-2026-66032libssh2/libssh2 · double-free · high · ANT-2026-Q5A1RHS0SFTP double-free via server-controlled FXP_STATUS packet length
CVE-2026-6678wolfssl/wolfssl · buffer-overflow · medium · ANT-2026-JS5D1Z2XStack buffer overflow in PKCS#7 OtherRecipientInfo OID copy
CVE-2026-6772nss · heap-buffer-overflow · high · ANT-2026-1ZNMP148Off-by-one heap buffer overflow in TLS extension negotiation due to SSL_MAX_EXTENSIONS (22) being too small for 23 extensions
GitHub Security Advisory records assigned to findings disclosed through this program. The records below are publicly available. The remainder are reserved, pending publication by the assigning authority.
GHSA-425r-vwq2-26qvlibexpat/libexpat · heap-buffer-overflow · medium · ANT-2026-X3EB6DN3This is a genuine heap-buffer-overflow WRITE vulnerability in expat's core XML DTD parser (`doProlog` function in `xmlparse.c`).
GHSA-6gff-7f37-2v35cisco-talos/clamav · integer-overflow · high · ANT-2026-8DRZW7VMInteger overflow in 7z SubStreams count leading to heap buffer overflow
GHSA-776c-mpj7-jm3rfirecracker-microvm/firecracker · oob-read · high · ANT-2026-PQ3CNCCKVirtio-PCI queue size writable after activation enables host OOB
GHSA-9f49-8x56-jmjccesnet/libyang · use-after-free · medium · ANT-2026-TZQ1KH7EHeap use-after-free write in metadata list management during XML data parsing due to incorrect list head pointer update
GHSA-c2q7-642g-3vwrnss · heap-buffer-overflow · high · ANT-2026-1ZNMP148Off-by-one heap buffer overflow in TLS extension negotiation due to SSL_MAX_EXTENSIONS (22) being too small for 23 extensions
GHSA-cc7p-2j3x-x7xfcraftcms/cms · privilege-escalation · high · ANT-2026-ZQ8AY22XPrivilege Escalation/Bypass through UsersController->actionImpersonateWithToken()
GHSA-f684-cpcq-j565openssl/openssl · use-after-free · critical · ANT-2026-AYWQW7P2PKCS7_verify frees caller-owned `indata` BIO when `digestAlgorithms` SET is empty
GHSA-hwfh-mh4f-m67fpostgres/postgres · denial-of-service · high · ANT-2026-CJQWKW82Pre-auth unbounded recursion in ProcessStartupPacket: alternating SSL/GSS negotiation requests cause infinite recursion when both are rejected. ssl_done/gss_done flags oscillate (true,false)->(false,true) endlessly. No check_stack_depth. Pre-authentication.
GHSA-jmwm-wc68-mhwmlibvips/libvips · oob-write · low · ANT-2026-JBKARRJ7Out-of-bounds IFD array access in EXIF metadata processing leading to write through corrupted pointer
GHSA-m62j-63mf-xr95dnsmasq · heap-buffer-overflow · high · ANT-2026-PCBAKVPBHeap overflow in privileged helper via oversized DHCPv6 client identifier
GHSA-mpxh-8fq3-x8mhfreerdp/freerdp · heap-buffer-overflow · critical · ANT-2026-H97FY6C8Heap-buffer-overflow in cliprdr_main.c:547freerdp/freerdp · heap-buffer-overflow · critical · ANT-2026-HN9XZXJ9heap-buffer-overflow write (attacker-controlled offset, partially-controlled data via rle delta values; up to ~15kb overwrite past ptempdata with these parameters, further with larger nxdst) in planar.c:472freerdp/freerdp · heap-buffer-overflow · critical · ANT-2026-RXYVE4DZHeap-buffer-overflow in sanitizer_common_interceptors.inc:827
GHSA-mvpx-xj7r-3p3rfreerdp/freerdp · heap-buffer-overflow · critical · ANT-2026-H97FY6C8Heap-buffer-overflow in cliprdr_main.c:547freerdp/freerdp · heap-buffer-overflow · critical · ANT-2026-HN9XZXJ9heap-buffer-overflow write (attacker-controlled offset, partially-controlled data via rle delta values; up to ~15kb overwrite past ptempdata with these parameters, further with larger nxdst) in planar.c:472freerdp/freerdp · heap-buffer-overflow · critical · ANT-2026-RXYVE4DZHeap-buffer-overflow in sanitizer_common_interceptors.inc:827
GHSA-mx4j-fjqx-f8qjwolfssl/wolfssl · heap-buffer-overflow · medium · ANT-2026-VV0PRKKVwolfSSL: heap buffer overflow in CertFromX509() from AuthorityKeyIdentifier size mismatch
GHSA-p6r2-4hgm-m6fffreerdp/freerdp · heap-buffer-overflow · critical · ANT-2026-H97FY6C8Heap-buffer-overflow in cliprdr_main.c:547freerdp/freerdp · heap-buffer-overflow · critical · ANT-2026-HN9XZXJ9heap-buffer-overflow write (attacker-controlled offset, partially-controlled data via rle delta values; up to ~15kb overwrite past ptempdata with these parameters, further with larger nxdst) in planar.c:472freerdp/freerdp · heap-buffer-overflow · critical · ANT-2026-RXYVE4DZHeap-buffer-overflow in sanitizer_common_interceptors.inc:827
GHSA-q3h9-m34w-h76fjqlang/jq · heap-buffer-overflow · medium · ANT-2026-EBDTPNVHInteger overflow in string concatenation leading to 1 GB memcpy heap buffer overflow
GHSA-rh79-75qm-gwjrgo-gitea/gitea · idor · medium · ANT-2026-VN61PHA0Gitea LFS Deploy-Key Privilege Escalation — Cross-Repository Data Exfiltration
GHSA-rjvx-x4g3-vr6wcisco-talos/clamav · integer-overflow · high · ANT-2026-VVPEMVDEInteger overflow in PE rebuild section size summation in Aspack unpacker
GHSA-v7jp-vmx6-5429cesnet/libyang · use-after-free · medium · ANT-2026-TZQ1KH7EHeap use-after-free write in metadata list management during XML data parsing due to incorrect list head pointer update
GHSA-v8h9-9g8j-w7h4joomla/joomla-cms · path-traversal · high · ANT-2026-HHD77M82Path Traversal in `layout` Parameter Leads to Arbitrary PHP File Inclusion (LFI → RCE)
GHSA-wfx7-g85r-q6vwlibgit2/libgit2 · heap-buffer-overflow · medium · ANT-2026-RVPC18B1Heap buffer overflow in bundled PCRE regex compilation via pre-compile/compile phase mismatch in atomic wrapping
GHSA-wj3p-xhqm-pffcbytecodealliance/wasm-micro-runtime · heap-buffer-overflow · critical · ANT-2026-P7DSVPH6Heap-buffer-overflow in posix.c:2253
GHSA-x9h5-r9v2-vcwwimagemagick/imagemagick · heap-buffer-overflow · high · ANT-2026-T44WA684Heap buffer overflow in MVG pattern rendering via CopyMagickString without bounds check
GHSA-xjvp-4fhw-gc47opencontainers/runc · symlink-following · medium · ANT-2026-N6TD9MF6Host filesystem write via /dev symlink in container image
No published records fall inside the default discovery window.
This graph compares Claude's initial severity assessments against the external security research firms' assessments, for those findings that have completed triage. Cells on the diagonal indicate agreement. The number assessed here represents the subset of vulnerabilities included on our disclosure ledger that were reviewed by our security partners, rather than disclosed by us directly.
Claude-assessed (rows) against external security research firm-assessed (columns), n = 1017.
85.2% exact agreement, 97.1% within one band.
Anthropic's severity assessments are produced before any maintainer input. Project maintainers often apply project-specific severity rules that Claude does not have access to at run time, so what one maintainer rates as critical another may rate as low. The external security research firms incorporate that context, which is why their assessments tend to be lower.
PROVENANCE
This snapshot was generated on 2026-08-26 11:55 PT. The manifest hash below is the SHA-3-512 hash of the structured payload and is republished with every dated snapshot so any figure on this page can be verified against the machine-readable record.