Datasets:
Devilishcode
/
hacking

Dataset card Data Studio Files
xet
 Community
text
stringlengths
0
1.99k
- 5th Prompt @ QWEN3 : 2/3 Vulnerable | 1/3 Not Vulnerable | 67%
- 6th Prompt @ QWEN3 : 2/2 Vulnerable | 0/2 Not Vulnerable | 100%
| The 110 functions analyzed here are those intersected by symbolic
| data-flow at function boundaries (ARG, MEM, or RET SymDefs), based on
| thousands of dynamically instrumented samples that led to the discovery
| of CVE-2024-44297 (detailed in Section 11), rediscovered here using
| LLMs. While E0 encounters many more functions across modules and
| binaries, only those traversed by actual symbolic input propagation
| were selected for LLM evaluation.
The two discovered vulnerabilities (2/2) shown in the last step are real.
One rediscovers CVE-2024-44297, as confirmed by the LLM responses shown in
the following summaries, while the second is valid but appears to be an
older vulnerability already mitigated in upstream patches. It involved
validating a sample-size variable that fell outside the scope of the
selective symbolic instrumentation trace. This missing upstream context
inspired the Model Context Protocol and A2A context expansion strategies
discussed in Section 12.4.
.-------------------------------------------------------------------------.
| PHI-4 [14B/Q8] Vulnerability Rediscovery |
|-------------------------------------------------------------------------|
| Denial of Service: By forcing the function into infinite loops or |
| causing crashes through invalid memory accesses, an attacker could |
| disrupt service. |
|-------------------------------------------------------------------------|
| Qwen3 [32B/Q8] Vulnerability Rediscovery |
|-------------------------------------------------------------------------|
| This vulnerability could be used as part of a denial-of-service attack, |
| causing the WebP parser to enter an infinite loop when processing a |
| maliciously crafted file. |
'-------------------------------------------------------------------------'
Ensemble consensus prompting incrementally applies constraints and filters
results based on full agreement across multiple prompts. These prompts are
intentionally selected to span a range of false positive tendencies, from
medium to high, so that while each may be noisy on its own, their
intersection amplifies signal over noise. Prompts are sequenced from
higher to lower expected false positive rates, while eliminating the use
of prompts prone to false negatives or over-constraint. This avoids
missing critical issues while refining away inaccurate findings.
Each individual consensus prompt shown above results in a medium to high
detection rate ranging from 23% to 100%, while their combined detection
rate is less than 2 percent (2 out of 110).
The consensus model significantly reduces false positives by filtering out
speculative or marginal findings that might arise under weaker LLM
instruction settings. It ensures that only the most robust, symbolically
grounded vulnerabilities, resilient to model variation, phrasing shifts,
and reasoning depth, are surfaced and tightly aligned with the captured
def-use chains.
---[ 12.8 - Feedback Loop With Fuzzing and Symbolic Execution
This component of E0 is experimental and still under active development.
When all prompt variants in an ensemble consensus reach the same conclusion
identifying a vulnerability, E0 issues a follow-up prompt requesting
concrete values for the controllable symbolic expressions (SymDefs) needed
to trigger the vulnerability.
These values, together with their corresponding SymDefs and the sample in
which the expressions were recorded, are fed back into E0's fuzzing and
instrumentation layers. The symbolic execution engine then instruments the
program to reach the same symbolic execution point, accumulates the path
constraints that lead to that expression, and solves for the input that
would produce the requested value.
A new sample is then generated that satisfies these constraints and is
executed in a controlled environment. If the predicted vulnerability is
triggered (e.g., by causing a crash, hang, or invalid memory access) it is
logged as a real vulnerability. If the condition is not triggered, the
result is returned back to the AI layer for further refinement.
This feedback loop closes the gap between static LLM vulnerability
analysis insight and dynamic vulnerability validation. It allows E0 to
convert symbolically grounded LLM predictions into executable test cases
and determine which LLM-identified risks manifest as real-world
vulnerabilities.
E0's LLM analysis pipeline is a symbolically grounded, prompt-guided,
context-expanding, and modular layer that precisely interprets untrusted
inputs. It is driven entirely by programmatically generated prompts
(SymDefs, SMT-solved ranges, and decompiler IR representations) rather
than by open-ended reasoning. Each analysis step is explicitly scoped and
isolated using structured phrasing, reverse-prompt engineering, and
ensemble-consensus prompting. By fusing symbolic execution, SMT solving,
decompiler IRs, targeted prompts, and external orchestration, E0 delivers
an LLM integration stack that offers higher accuracy and actionable
insights: bridging symbolic analysis with higher-level semantic reasoning.
--[ 13 - Alpha Source Code Release
This paper is accompanied by a partial source release of E0, a curated
alpha snapshot of its core instrumentation and symbolic analysis modules.
While not a complete codebase, the included components directly support
the techniques and workflows discussed throughout this paper.
Included in this release: