Friday, 20 November 2026
2 agent hacks today▼ 9 vs yesterday (11)

Three critical flaws disclosed in Flyto2 Core agent runtime

GitHub advisories published on 20 November 2026 describe an arbitrary file write and two ways around the workflow engine's anti-SSRF guard.

By The Agentic Times ·

A humanoid machine with exposed circuits stands in an open vault, its metal hands writing files onto unguarded shelves while invisible barriers crumble around it.
A humanoid machine with exposed circuits stands in an open vault, its metal hands writing files onto unguarded shelves while invisible barriers crumble around it. · Illustration: The Agentic Times

Three security advisories published on GitHub on 20 November 2026 describe vulnerabilities in Flyto2 Core, the Python package flyto-core that runs Flyto's recipe and module workflows. The most serious, tracked as GHSA-p34x-fmph-9fjx and CVE-2026-77312, allows an attacker to write a file of their choosing to any path on the host. The advisory rates it 9.3, critical, under CVSS v3.1. The other two, CVE-2026-77315 and CVE-2026-77316, both defeat the product's protection against server-side request forgery, in which a server is tricked into fetching an internal address on an attacker's behalf.

According to the first advisory, the file write stems from an incomplete fix. An earlier critical arbitrary-write issue, GHSA-2956-977x-2w3r, was patched in version 2.26.6 by routing a hand-picked list of modules through a path-checking function called validate_path_with_env_config(). That list did not include the data.* and file.* write families. The advisory says data.csv.write checks only for the substring .. in the supplied path, a denylist that absolute paths do not trip, while data.json_to_csv, file.copy and file.move perform no path check at all. There is no central enforcement in the runtime, the advisory states, so any module that omits the per-module call is unconfined.

The practical effect, according to the advisory, is that a client able to run a module or recipe can write attacker-controlled content outside the directory named in FLYTO_SANDBOX_DIR. Overwriting a Python module in the install directory, a .flyto recipe, a shell startup file, a crontab or ~/.ssh/authorized_keys would give code execution on the host. A published proof of concept shows the sandbox guard rejecting an out-of-sandbox path while the unguarded modules write to that same path successfully.

The second advisory, GHSA-2mr3-rxrq-238c, concerns the http.paginate module. It validates only the first URL supplied by the caller, then follows pagination targets handed back by the remote server, such as Link: rel="next" headers or cursor URLs, without revalidating them. The advisory says the follow loop contains no guard calls. In a live test, a request with an allowlisted base address of 127.0.0.1 followed a Link header pointing at a non-allowlisted loopback address and returned the internal marker string in its aggregated results. The suggested fix is to run every follow-up URL through the existing validator before fetching it. The advisory lists versions up to and including 2.26.6 as affected.

The third, GHSA-6pm8-6f34-9v3g, is rated 6.8, medium. It describes a DNS rebinding weakness: validate_url_ssrf resolves a hostname, checks the returned addresses, then passes the hostname string on to the HTTP client, which resolves it again independently at connect time. An attacker controlling DNS with a zero time-to-live can answer with a public address for the check and a private or cloud metadata address for the connection. The advisory says every guarded HTTP consumer is affected, including http.get, graphql.query, the Slack, Discord and Teams notification modules and ai.vision_analyze, and that earlier SSRF hardening work, including redirect revalidation, does not help because the first hop already rebinds. The recommended remediation is to pin the validated IP address and connect to it directly.

All three advisories list the affected component at HEAD ae2b38e, on the v2.26.7 line. No fixed version is named in the advisories.

Sources

  1. github.comhttps://github.com/advisories/GHSA-p34x-fmph-9fjx
  2. github.comhttps://github.com/advisories/GHSA-2mr3-rxrq-238c
  3. github.comhttps://github.com/advisories/GHSA-6pm8-6f34-9v3g