Skip to content

Commit 3c12146

Browse files
rng1995claude
andcommitted
refactor(taint): narrow interprocedural taint to bounded direct calls
Two review rounds of the broad interprocedural design showed it kept producing denial-of-service paths (per-call work proportional to the number of candidate classes, unbounded receiver class lists, per-source fact sets, messages copying long identifiers) and recall or precision regressions (receiver typing that hid real callees, a fallback that bound into every same-named method, callback rules that bound the wrong arguments). This narrows interprocedural flow to calls whose callee is known from syntax alone, in O(1) per call. Kept: - Lexical scopes. Function locals and parameters are per function and module globals keep bare keys; global/nonlocal and closures follow Python's rules; class bodies are invisible to methods; lambdas and comprehensions are scopes, with PEP 572 walrus binding. Names resolve in one depth-first pass over the scope tree with a stack per identifier, O(1) per name however deep the nesting. Keys are small ints built from a scope id and a reference to the identifier, never from a qualified name. - Direct calls by name to functions defined in the file (including nested ones visible lexically and lambdas assigned to a name): positional and keyword arguments, defaults, *args/**kwargs forwarding, 16 positional slots. Every definition bound to one name is one callee, so binding work is linear in calls, not calls x redefinitions. - Per-call-site returns. A callee's return summary only carries facts that do not depend on its parameters; what a call returns from its own arguments is read at that call site from those arguments (a call's value includes its arguments, as before). One caller's secret no longer taints other callers' results, also through closures and lambdas. - Methods, only where the callee is syntactic: self.m()/cls.m() bind to the class's own method or its nearest in-file base (a precomputed lookup over at most 8 classes), Class.m(obj, ...) with the unbound offset, super().m(), and Class(...)/cls(...) to __init__. Aliases of a class bind nothing, as on main. self.x/cls.x attributes are per class: a read sees stores in its class and in-file bases, never in siblings. - One source per key and sink lane, the most severe for the lane: credential > file read > external input for network sinks, and external input > file read > credential for code execution and deserialization sinks. A helper's source can no longer turn a TT3 into TT2/TT4, nor a TT5/TT6 into TT2. When the lanes agree (almost always) they share one fact, which propagates once. - Two-phase propagation. Phase A replays main's assignment flows in main's order; phase B adds argument binding, returns, attributes and defaults, and only fills clean keys or upgrades to a more severe source. A flow main reports keeps main's message, except where a more severe source now reaches the same variable. - Messages name what the sink spells (never an attribute spelling from elsewhere in the file); identifiers longer than 120 characters are shortened, so message bytes stay bounded per finding. Dropped, each a source of DoS paths or regressions: - receiver type inference (per-scope type maps, annotations, with-as, cast and builtin results), MRO linearization beyond the bounded base lookup, and @Property reads; - the untyped-receiver fallback into every same-named method; - callback rules (Thread target, submit, run_in_executor): as on main; - class hierarchy families and the shared-descendant attribute rule; - per-source fact sets; - comprehension iteration flows (main does not model iteration targets, and they changed which variable messages cite at sinks); - lambda return summaries (a lambda's value is already read by name). Bounds: indexing, name resolution, flow building, callee linking and propagation are linear, and check_runtime runs in each of their loops. Every flow fires at most once per (lane, rank, slot), 12 times whatever the number of sources. On every DoS repro from both reviews, every file completes and later files keep their findings; time is 1.0-2.2x main and peak memory matches main, except +14% on a variable carrying all 23 sources. Tests: - main's 77 taint tests pass unchanged. - New: name-clash classes found in real skills, direct-call recall, methods and attributes without sibling leakage, source lanes, message compatibility, deadlines in indexing, flow building, callee linking and propagation, and every DoS repro from both reviews as a work-scaling test (executed analyzer lines at 1x and 4x size, not wall time), a fact count or a byte count. - TestNotResolved pins main-like behavior for the dropped features. - Removed PR tests, with the reason: - receiver typing and the untyped fallback (library instance elsewhere, untyped parameter receiver, typed receivers reading or binding only their class, returns of constructed and attribute receivers, and their exec/file variants): the features are dropped; instance receivers now bind nothing, as on main; - callbacks on attribute, constructor and parameter receivers, Thread kwargs, Thread(target=Base.m): callback rules are dropped; - @Property reads, Config.TOKEN read through the class name, mixins and subclass overrides reached from a base: dropped with hierarchy families; overrides now assert main-like results; - comprehension over tainted data tainting its target: iteration flows are dropped; - the 1 s workflow-deadline wall-clock test: replaced by operation counts (it also passed on the unfixed code). The other PR tests were kept under new names (scope precision, helper returns, argument binding, closures, attributes, super, unbound calls, context sensitivity, lambda and comprehension scopes, deadlines). Signed-off-by: Narendran Raghavan <[email protected]> Co-Authored-By: Claude Opus 5.5 <[email protected]>
1 parent 033b4cf commit 3c12146

2 files changed

Lines changed: 2410 additions & 2488 deletions

File tree

0 commit comments

Comments
 (0)