CVE-2026-16760: fickling: safe-import trust bypass allows pickle RCE

In Trail of Bits fickling through 0.1.12, the _process_import() method in fickle.py adds an imported name to the likely_safe_imports set whenever the source module is present in sys.stdlib_module_names and the name is absent from the UNSAFE_IMPORTS denylist, without inspecting what the name resolves to or how it behaves. The OvertlyBadEvals analysis pass in analysis.py then unconditionally skips any call whose function name appears in likely_safe_imports, including calls that would otherwise be flagged as overtly dangerous. As a result, any callable reachable through a non-blocked standard library module inherits module-level trust regardless of its own behavior, and check_safety() returns LIKELY_SAFE with zero findings for pickle payloads that reach code execution or file write through such callables. Demonstrated primitives: dataclasses._create_fn, which passes an attacker-controlled function body to exec() and yields arbitrary code execution on Python 3.7 through 3.12 (the internal _create_fn function was removed in Python 3.13); mailcap.findmatch, which internally calls os.system() on a test command taken from an attacker-controlled caps dictionary, yielding arbitrary code execution on Python versions through 3.12 (the mailcap module was removed in Python 3.13); and readline.add_history() followed by readline.write_history_file(), which writes attacker-controlled content to an arbitrary filesystem path (POSIX platforms, where the readline module is available). fickling.load() chains check_safety() into pickle.loads() as an explicit security gate, so a LIKELY_SAFE verdict causes the payload to be deserialized and executed with the privileges of the victim process. The root cause is the inference of callable safety from module origin; it is independent of, and separately fixable from, the dotted __qualname__ resolution weakness tracked as CVE-2026-118938, which this advisory does not address. As of 2026-10-30 the three demonstrated modules were added to UNSAFE_IMPORTS on the vendor master branch (commit d011709) pending release; released versions through 0.1.12 remain affected, and the module-origin trust root cause persists, as demonstrated by the related CVE-2026-118938 remaining exploitable on that same build.

An attacker publishes a model checkpoint on a public hub or sends it to a data science team as a shared artifact. The pickle contains a crafted call to dataclasses._create_fn with an attacker-controlled function body, or mailcap.findmatch with a malicious test command. The team's ingestion pipeline runs fickling.load(), which returns LIKELY_SAFE with no findings and then deserializes the file. The payload runs inside the training or inference host, steals cloud credentials and API keys, and installs persistence. Alternatively, the readline primitive overwrites ~/.bashrc or an authorized_keys file to gain a later foothold.