Reject malformed ELF symbol metadata

[zardus]

THIS MESSAGE WAS GENERATED BY AN AUTOMATED PROCESS

Problem

Three unavailable ARM Cortex-M ELF binaries carry relocations whose symbol index is 2,097,161 even though their section-backed symbol tables contain only 553 or 662 entries. Loading each binary logs an ELFParseError: expected 4, found 0 traceback while pyelftools tries to read beyond the table:

• 32648bd04baf266713d8f5b26a7d0279bd0d34ffa339101e08929b1454ba6c97
• 8724d18dfbb82d33eada4f66875b846ff68026dc8c15b0505b8bb1a524ba1915
• 95cd4fbe94fc27af342b9de952615d6061c704f6b43a39fb131b5180ff7d1a3d

The failure is observable while ELF.__register_relocs resolves relocation symbol index 0x200009 in each binary.

Malformed GNU version metadata has the same trust-boundary problem: a version-table ordinal absent from the parsed version definitions raises KeyError while constructing a symbol.

Root cause

ELF.get_symbol passes integer symbol identifiers directly to pyelftools. A section-backed SymbolTableSection.get_symbol computes an entry offset without checking num_symbols(), so an untrusted relocation index can seek beyond the section. The synthetic table inferred from PT_DYNAMIC is different: it has no section type, and its inferred count is not a reliable upper bound.

The version lookup similarly indexed _versions under the assumption that every ordinal referenced by .gnu.version has a matching definition.

Fix

Reject negative indices and indices at or beyond num_symbols() for real SHT_SYMTAB and SHT_DYNSYM tables, returning None so the malformed relocation is skipped. Preserve the existing synthetic PT_DYNAMIC behavior, and use a non-raising lookup for unknown version ordinals.

Focused tests cover negative and oversized indices, an unknown version ordinal, the synthetic-table exception, and the PLT/extern paths that depend on that exception.

Testing

The affected objects were measured before and after with one store-built environment per side. OBJECT denotes each unavailable binary identified by the corresponding SHA-256 above.

ENV=/nix/store/6n7araxwls4qn22g82qvjzwr4iwd1sly-python3-3.12.13-env; OBJECT='<unavailable-binary>'; RTDB_BASE=$(mktemp -d); export OBJECT RTDB_BASE; "$ENV/bin/python" -P -c 'import angr,logging,os; rows=[]; h=logging.Handler(logging.ERROR); h.emit=rows.append; log=logging.getLogger("cle.backends.elf.elf"); log.propagate=False; log.addHandler(h); p=angr.Project(os.environ["OBJECT"],auto_load_libs=False); log.removeHandler(h); o=p.loader.main_object; print(type(o).__name__,o.arch.name,len(o.symbols),len(o.relocs),len(rows),sum(r.exc_info is not None and type(r.exc_info[1]).__name__=="ELFParseError" for r in rows))'

ENV=/nix/store/2mijvwvc1dl632dxn4s09cfpykzqavkx-python3-3.12.13-env; OBJECT='<unavailable-binary>'; RTDB_BASE=$(mktemp -d); export OBJECT RTDB_BASE; "$ENV/bin/python" -P -c 'import angr,logging,os; rows=[]; h=logging.Handler(logging.ERROR); h.emit=rows.append; log=logging.getLogger("cle.backends.elf.elf"); log.propagate=False; log.addHandler(h); p=angr.Project(os.environ["OBJECT"],auto_load_libs=False); log.removeHandler(h); o=p.loader.main_object; print(type(o).__name__,o.arch.name,len(o.symbols),len(o.relocs),len(rows),sum(r.exc_info is not None and type(r.exc_info[1]).__name__=="ELFParseError" for r in rows))'

All 3 affected objects changed from 3 cle ELF error records, including 1 ELFParseError, to 2 error records and 0 ELFParseErrors. Backend, architecture, entry point, address range, section and segment inventories, symbol count, and relocation count stayed identical.

Regression set: 3 other ARM Cortex-M ELF objects from the same decbench sweep family. Their complete probe summaries were byte-for-byte identical before and after; all loaded with 0 cle ELF error records, and 0 of 3 got worse.

Local validation at head 5218e54ebbb290693881fcf047a81a56e2fc65e3:

• focused ELF and PLT/extern regressions: 7 passed
• cle: 286 passed, 9 skipped; one unrelated missing-fixture failure from the already-merged Mach-O: Load a file that carries no LC_SYMTAB #843, whose prerequisite Add a Mach-O relocatable object that carries no symbol table binaries#235 is still open
• merge-base pylint: 10.00/10.00 for both changed files; pyright: 0 added errors
• full gate: test-inputs, test-packages, pre-commit, feature-build, archinfo, pypcode, pyvex, pysoot, angr Rust, and worktree cleanliness passed; unrelated existing failures remained in workspace campaign tests, the mono component-list invariant, the missing cle fixture above, and one unchanged angr reaching-definitions test

The complete cle suite requires the fixture from angr/binaries#235, which must merge before this pull request.

sync: angr/binaries#235

Validation: #853 (comment)

session: sharpen

[zardus]

THIS MESSAGE WAS GENERATED BY AN AUTOMATED PROCESS

Complete stdout from the loader-error probe for one affected unavailable binary; the fields are backend, architecture, symbol count, relocation count, cle ELF error count, and ELFParseError count.

Before — pyelftools reads beyond the section-backed symbol table and cle records an ELFParseError:

cle master 997d432

ELF ARMCortexM 662 0 3 1

After — the invalid index is rejected before parsing, so no ELFParseError is recorded:

with 5218e54

ELF ARMCortexM 662 0 2 0

[zardus]

THIS MESSAGE WAS GENERATED BY AN AUTOMATED PROCESS

Validation record for head 5218e54ebbb290693881fcf047a81a56e2fc65e3 against baseline 997d4322678fe5876d9f83e0dd34e1052813b9db.

Dataset A/B, using one store-built environment per side and each of the three unavailable binaries identified by SHA-256 in the pull request body:

ENV=/nix/store/6n7araxwls4qn22g82qvjzwr4iwd1sly-python3-3.12.13-env; OBJECT='<unavailable-binary>'; RTDB_BASE=$(mktemp -d); export OBJECT RTDB_BASE; "$ENV/bin/python" -P -c 'import angr,logging,os; rows=[]; h=logging.Handler(logging.ERROR); h.emit=rows.append; log=logging.getLogger("cle.backends.elf.elf"); log.propagate=False; log.addHandler(h); p=angr.Project(os.environ["OBJECT"],auto_load_libs=False); log.removeHandler(h); o=p.loader.main_object; print(type(o).__name__,o.arch.name,len(o.symbols),len(o.relocs),len(rows),sum(r.exc_info is not None and type(r.exc_info[1]).__name__=="ELFParseError" for r in rows))'

ENV=/nix/store/2mijvwvc1dl632dxn4s09cfpykzqavkx-python3-3.12.13-env; OBJECT='<unavailable-binary>'; RTDB_BASE=$(mktemp -d); export OBJECT RTDB_BASE; "$ENV/bin/python" -P -c 'import angr,logging,os; rows=[]; h=logging.Handler(logging.ERROR); h.emit=rows.append; log=logging.getLogger("cle.backends.elf.elf"); log.propagate=False; log.addHandler(h); p=angr.Project(os.environ["OBJECT"],auto_load_libs=False); log.removeHandler(h); o=p.loader.main_object; print(type(o).__name__,o.arch.name,len(o.symbols),len(o.relocs),len(rows),sum(r.exc_info is not None and type(r.exc_info[1]).__name__=="ELFParseError" for r in rows))'

Result: all 3 affected objects changed from 3 cle ELF error records with 1 ELFParseError to 2 error records with 0 ELFParseErrors. The 3-object same-family regression set remained byte-for-byte identical with 0 errors before and after; worse: 0.

Focused regression command:

cd features/fix-c-23/repos/cle && ../../../../nix/run.sh --feature fix-c-23 -- pytest --import-mode=append -q tests/test_elf.py tests/test_extern.py::test_ppc64_abiv1_untyped_function_import tests/test_plt.py::TestCheckPltEntries::test_x86_64_fauxware tests/test_plt.py::TestCheckPltEntries::test_x86_64_simple_overflow_nopie

Result: 7 passed in 0.58s.

Merge-base CI diff checks:

nix/run.sh --feature fix-c-23 -- python3 -P .agents/skills/angr-validate-workspace/scripts/run-ci-diff-checks.py --repository features/fix-c-23/repos/cle

Result: pylint 10.00 to 10.00 for cle/backends/elf/elf.py; new tests/test_elf.py 10.00. Pyright errors 46 to 46 for production and 0 to 0 for the test.

Complete collection gate:

./feature.sh test fix-c-23

Passed suites: test-inputs, test-packages, pre-commit, feature-build, archinfo (42 tests and 34 subtests), pypcode (46 tests and 187 subtests), pyvex (87), pysoot, angr Rust, and worktree cleanliness. Cle passed 286 tests and skipped 9; its only failure is a missing Mach-O fixture introduced by already-merged #843 and still waiting on angr/binaries#235. Angr passed 3,747 tests, skipped 47, xfailed 2, and failed one unchanged test that passes a now-invalid third argument to Register. Workspace campaign tests and mono each have unrelated existing failures. The feature did not adopt angr-management, so that suite did not run.

[angr-bot]

Corpus decompilation diffs can be found at angr/dec-snapshots@master...angr/cle_853