Loading vulnerability…
CVE-2026-72742 is a file exfiltration vulnerability in Stanford NLP DSPy affecting versions up to and including 3.3.0b1. The flaw is in the Image and Audio typed output handling path, where untrusted language model completions are parsed by JSONAdapter or ChatAdapter through parse_value into TypeAdapter validation. During that flow, a attacker-influenced value supplied in the url field of a parsed Image or Audio object can be treated as a local filesystem path. The vulnerable logic in the image and audio encoding paths checks whether the supplied value refers to a local file and, if so, reads and base64-encodes that file. The encoded contents are then incorporated into outbound prompt messages. This creates a path-based arbitrary local file read and exfiltration condition when applications process untrusted model output through these adapters.
Mallory correlates every CVE against your assets, your vendors, and active adversary campaigns. Know which vulnerabilities matter for you, not just which ones are loud.
What it means. What to do now. Patch path, mitigations, and the assume-compromise checklist.
What an attacker gets, and what they’ve been doing with it.
If you can’t patch tonight, do this now.
Patch, then assume compromise.
No public exploits tracked yet. Mallory keeps watching.
No public exploit code observed for this vulnerability.
2 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.
Query your assets running an affected version, and investigate the blast radius.
Every observed campaign linking this CVE to a named adversary.
Malware families riding this exploit, with evidence and IOCs.
YARA, Sigma, Snort, and vendor rules, auto-deployed to your SIEM.
Cross-references every affected SKU, including bundled OEM variants.
Community discussion across Reddit, Mastodon, and other social sources.