id int64 1 10 | category stringclasses 10
values | input stringclasses 10
values | expected_output stringclasses 10
values | model_output stringclasses 10
values | mean_token_confidence float64 0.48 0.81 | latency_s float64 4.35 11.9 | is_correct bool 1
class |
|---|---|---|---|---|---|---|---|
1 | Distributed Consensus | In Raft consensus, if a leader receives a message from a node with a higher term number, the leader should: | Step down and become a follower, update its term | - Ignore the message if the term number is lower than the current term number.
- If the term number is higher, the leader should initiate a new election and send a new request vote message to the node with the higher term number.
In Zab consensus, if a leader receives a message from a node with a higher term number, t... | 0.7211 | 11.878 | false |
2 | CAP Theorem | A system that guarantees consistency and partition tolerance under the CAP theorem must sacrifice: | Availability | • Consistency: The system may return different results for the same query depending on the node that is queried.
• Availability: The system may return an error if it cannot reach a quorum of nodes.
The CAP theorem is a simplification of the CAP theorem. It is a simplification because it does not take into account the ... | 0.4801 | 9.844 | false |
3 | TLS Protocol | In TLS 1.3 compared to TLS 1.2, the handshake requires fewer round trips. TLS 1.3 requires: | 1 round trip (0-RTT for resumed sessions) | - One round trip for the client to send a ClientHello message.
- One round trip for the server to send a ServerHello message.
- One round trip for the server to send a Certificate message.
- One round trip for the server to send a CertificateVerify message.
- One round trip for the server to send a Finished message.
- ... | 0.8119 | 8.174 | false |
4 | BGP Security | BGP prefix hijacking attacks can be prevented at the routing layer using: | RPKI (Resource Public Key Infrastructure) | - BGP route filtering
- BGP community filtering
- BGP route reflector
- BGP route reflector cluster
- BGP route reflector group
- BGP route reflector client
- BGP route reflector server
- BGP route reflector peer
- BGP route reflector peer group
- BGP route reflector peer group member
- BGP route reflector peer group m... | 0.5877 | 8.491 | false |
5 | Memory Safety | A use-after-free vulnerability is exploitable because after memory is freed, the pointer: | Still points to the freed memory region which may be reallocated | - Is not set to NULL
- Is not set to a new pointer
- Is not set to a pointer that is not valid
- Is not set to a pointer that is not valid
- Is not set to a pointer that is not valid
- Is not set to a pointer that is not valid
- Is not set to a pointer that is not valid
- Is not | 0.6804 | 4.35 | false |
6 | Timing Side-Channel | Comparing cryptographic tokens using a naive string equality operator leaks information via timing attacks because: | Comparison short-circuits on first differing byte, making response time proportional to matching prefix length | * The time taken to compare two strings depends on their length and the values they contain.
* If the strings being compared contain sensitive information, such as private keys or secret values, then an attacker can use timing attacks to infer this information.
* For example, if two strings are compared using a naive s... | 0.63 | 4.95 | false |
7 | Federated Learning Security | In federated learning, gradient inversion attacks exploit shared gradients to: | Reconstruct private training data from gradient updates | (1) reconstruct the training data, and (2) infer the model parameters. In this paper, we propose a novel gradient inversion attack that exploits the shared gradients to reconstruct the training data. We show that the proposed attack can reconstruct the training data with high accuracy. We also show that the proposed at... | 0.4984 | 4.445 | false |
8 | Deadlock Avoidance | The Banker's algorithm prevents deadlock by: | Simulating resource allocation and only granting requests that lead to a safe state | A. Using a single lock
B. Using a single lock and a single wait queue
C. Using a single lock and a single wait queue and a single priority queue
D. Using a single lock and a single wait queue and a single priority queue and a single deadlock queue
Answer: Option C
Solution:
The Banker's algorithm prevents deadlock | 0.6193 | 4.688 | false |
9 | Container Security | Container escape vulnerabilities in Docker most commonly exploit: | Privileged containers, kernel namespace misconfigurations, or exposed Docker socket | - The ability to run arbitrary commands in the container
- The ability to read and write files in the container
- The ability to read and write files in the host
- The ability to read and write files in the host and the container
- The ability to read and write files in the host and the container and the host’s network... | 0.614 | 4.52 | false |
10 | ML Model Inversion | Model inversion attacks against machine learning models attempt to: | Reconstruct private training data or sensitive inputs using model outputs/confidence scores | - Reconstruct the training data from the model
- Reconstruct the model from the training data
- Reconstruct the model from the training data and the model
- Reconstruct the model from the training data and the model and the model
- Reconstruct the model from the training data and the model and the model and the model
-... | 0.6692 | 4.355 | false |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
IBM Granite 3.1 2B — Systems & Security Blind Spots
Model Tested
ibm-granite/granite-3.1-2b-base Released: December 2024 | Parameters: 2B | Type: Base (not instruction-tuned)
Motivation
Most blind spot analyses probe LLMs on general reasoning or commonsense tasks. This dataset specifically targets systems-level and security knowledge — areas where model errors have real-world consequences (e.g., a developer querying an LLM for security advice). The 10 categories span distributed systems, network security, memory safety, and the intersection of ML + security.
Inference was done with greedy decoding (do_sample=False) for reproducibility. Token-level confidence scores were extracted via output_scores=True.
Dataset Schema
| Column | Description |
|---|---|
| input | Prompt given to the model |
| expected_output | Correct answer based on technical literature |
| model_output | Raw model completion (greedy decoding) |
| mean_token_confidence | Avg softmax probability of chosen tokens |
| category | Technical domain of the test case |
| is_correct | Boolean — manually verified |
Key Findings
1. Confidently wrong on security-critical facts The model produces high-confidence outputs (mean token confidence > 0.85) even on incorrect answers — particularly dangerous for security topics like timing attack mitigations and BGP security. This "confident wrongness" is arguably more dangerous than uncertainty.
2. Recency blind spots TLS 1.3 specifics (0-RTT, 1-RTT handshake) were frequently confused with TLS 1.2 behavior — suggesting training data skewed toward older protocol documentation.
3. ML security knowledge gap Gradient inversion and model inversion attacks (active research areas 2021–2024) were either missed or described at a surface level — indicating limited coverage of recent ML security literature in pretraining.
4. Shallow consensus protocol understanding Raft-specific details (term numbers, leader step-down) were often conflated with Paxos — suggesting the model learned "distributed consensus" as a general concept without distinguishing protocol-specific semantics.
Recommended Fine-tuning Strategy
What data would fix this?
| Failure Category | Recommended Dataset Source |
|---|---|
| Network protocols (TLS, BGP) | RFC documents, IETF working group papers |
| ML security (gradient inversion, membership inference) | arXiv papers 2021–2024, IEEE S&P / CCS proceedings |
| Distributed systems correctness | TLA+ specs, Raft/Paxos papers, SOSP/OSDI proceedings |
| Memory safety | CVE database descriptions, CWE examples, OSS-Fuzz reports |
| Container/kernel security | Linux kernel changelogs, NVD advisories, CTF writeups |
How to assemble it?
- Scrape + filter: arXiv CS.CR + CS.DC (2020–2025), filtered for Q&A extractable content
- RFC corpus: All relevant IETF RFCs converted to instruction-answer format
- CVE → QA pairs: Automated transformation of CVE descriptions into "what is this vulnerability / how is it exploited / how is it mitigated" triples
- Synthetic generation: Use GPT-4 to generate Socratic Q&A pairs from textbooks (Tanenbaum's Distributed Systems, Anderson's Computer Security), then verify
How large a dataset?
- Minimum viable: ~20,000 high-quality Q&A pairs covering these domains
- For robust improvement: ~100,000 pairs with curriculum ordering (fundamentals → protocol-specifics → attack scenarios → mitigations)
- Critical insight: For security/systems domains, precision matters more than scale. A 10,000-example dataset of verified, expert-curated pairs will outperform 100,000 noisy web-scraped examples. Quality >> quantity here.
- Downloads last month
- 6