Datasets:
Devilishcode
/
hacking

Dataset card Data Studio Files
xet
 Community
text
stringlengths
0
1.99k
This strategy allows E0 to expand context dynamically without overwhelming
the LLM or wasting token budget on irrelevant code. It also improves
modularity and parallelism, enabling scalable analysis through a set of
lightweight, focused AI agents, each focused on a narrow, verifiable
reasoning task.
---[ 12.5 - Modular Reasoning vs Chain-of-Thought Prompts
Although Chain-of-Thought (CoT) prompting is often promoted for complex
reasoning tasks, we found that it introduces significant false positives
and negatives in the context of vulnerability analysis. When multiple
reasoning steps are embedded into a single prompt, LLMs often produce
contradictions, overlook symbolic constraints, or fixate on irrelevant
code paths, resulting in less reliable analysis.
E0 avoids this by decomposing the reasoning process into modular,
independently controlled steps. Each step is executed within a separate
inference context, scoped to a narrow analysis goal. These include
evaluating a specific symbolic expression, analyzing a function-level
security pattern, or interpreting a solved constraint range. The results
are then aggregated externally, allowing the system to review, retry, or
refine intermediate steps as needed.
This modular approach improves accuracy, interpretability, and debugging.
It also allows the orchestration layer to selectively retry or expand any
single reasoning step without discarding the entire analysis. Rather than
embedding multi-stage logic into a single inference window, E0 treats each
step as a query-and-response unit, giving the orchestration layer full
control over the analysis flow.
While future reasoning models may support more reliable structured CoT
prompting and larger context windows, external modular reasoning remains
more robust and transparent in our current pipeline.
---[ 12.6 - Reverse Prompt Engineering
Reverse Prompt Engineering is a technique where the desired output is
provided to the model, which is then tasked with generating the most
accurate prompt to reproduce that output. In our workflow, this method
was adapted to address cases of false positives and negatives. We applied
the technique per model and per quantization level to improve prompt
clarity and effectiveness.
A key mechanism is instructing the model to critique its own prior output
by asking questions like:
| Which word or phrase in the original request
| led you to that incorrect assumption?
| Which word or phrase in the original request would
| you modify or remove to prevent this mistake?
In one real case, the symbolic expression *(uint64_t*)x10_1 was used in
the function sub_221204180. When the prompt instructed the model to
analyze its "value range", the model reported:
- Out-of-Bounds Memory Access
- Use-after-Free
- Arbitrary Memory Access
However, the value loaded from the dereferenced pointer was not used
in any memory access. Instead, it was treated as a plain operand in
arithmetic operations. The pointer itself was not attacker-controlled.
Only the value it pointed to was symbolized. When the prompt was revised
to ask about the "output range" rather than the "value range," the model
correctly responded:
| No direct vulnerabilities observed; expression
| is only used in a bitwise operation and shift.
This minor vocabulary change eliminated the false positive and produced a
more accurate analysis. Reverse Prompt Engineering revealed that the term
"value" had implicitly led the model to focus on pointer dereferencing
safety rather than the meaning of the loaded result.
This insight directly informed the vocabulary and framing of prompts used
across E0's ensemble prompting system.
---[ 12.7 - Ensemble Consensus Prompting
Ensemble Consensus Prompting is a technique in which multiple
independently phrased prompts are issued to analyze the same symbolic
slice or untrusted input. Instead of embedding many rules or constraints
into a single complex prompt, E0 distributes them across a sequence of
prompt variants. Each prompt introduces a focused subset of constraints
or directives, which improves model focus and reduces confusion.
A candidate vulnerability is elevated only when all prompt variants reach
the same conclusion, confirming the presence of a vulnerability. In
addition to strict consensus, E0 also tracks results of positive versus
negative findings across prompts to assist with manual debugging and
deeper analysis.
The following example illustrates how this process works:
- 1st Prompt @ PHI-4 : 25/110 Vulnerable | 85/110 Not Vulnerable | 23%
- 2nd Prompt @ PHI-4 : 13/25 Vulnerable | 12/25 Not Vulnerable | 52%
- 3rd Prompt @ PHI-4 : 8/13 Vulnerable | 5/13 Not Vulnerable | 61%
- 4th Prompt @ QWEN3 : 3/8 Vulnerable | 5/8 Not Vulnerable | 37%