bakakuk.com — Agentic AI hacker. Source code only. Built in Malaysia.

Agentic AI Hacker #MalaysiaBoleh

Bakakuk reads a repository the way a security researcher does: it forms concrete attack hypotheses, works to disprove each one against the source, and reports only what survives, with the cited lines and a reproduction a human can run. It never executes the target.

[ Fig. 0 ] zero-days,found and disproven from the codebuilt in malaysia

Breadth at low cost: one accepted finding from a $3 scan

The latest scan proposed 101 hypotheses, analyzed 54 of them, and ended with one finding the engine's own adversarial pass could not refute — for $3.01 and just over five hours of wall time. Here is the funnel and what is still owed before it counts as proven.

Announcement
Engine

What the new engine can and cannot move on the benchmark

The published scores measure a model reading a window of vulnerable code. They cannot see the hunt, so they cannot see this engine. Here is which changes could raise a full-hunt score, which cannot, what we corrected in the published runs, and the one case we retired.

Benchmark
Engine

The benchmark

Statistics

n-day benchmark recall:88% · engine 3.12.0

Scans completed:40

Hypotheses analyzed:1,006 of 2,488 proposed

Confirmed / candidate:13 confirmed · 28 candidate

Total model spend:$64.40

Credited CVEs:0

Numbers as of:2026-10-26

Featured(20)

Filter

Type

Topic

[ Fig. 3 ]10x

DateNameType

Get help(4)

Benchmark

Read how recall is measured before you read the score. Ground truth comes from real fix commits, and every run is stamped with the engine version.

Example: “Why is a low score published?”
Example: “What counts as found?”
Example: “Which run is the latest?”

Open the benchmark

Verdicts

Every hypothesis ends in one of six verdicts and most end in Refuted. The tiers say how much was proven, not how confident the model sounded.

Example: “What is the difference between Candidate and Confirmed?”
Example: “Why is an abstention never a refutation?”

Read the verdict tiers

Reports

An accepted finding is a report with a reproduction that cites only lines that exist in the repository and that a human can run by hand.

Example: “What does a reproduction look like?”
Example: “Why is nothing executed?”

How Bakakuk refutes itself

Disclosure

Found something in a project we published on? Mail [email protected] and we will route it to the maintainers with you.

Example: “How do I report a bug in a write-up?”
Example: “Can Bakakuk look at my repository?”

Contact us

Router(12)

Info

Four ways in. Each objective has a main link and three detours.

[ Fig. 42 ]10x

Topics

Scan a repository

Point Bakakuk at a GitHub repository with a time or dollar cap. It clones, profiles, mines the fix history, forms hypotheses and tries to break each one against the source.

Anatomy of one scan

Read a verdict

Six tiers from Refuted to Confirmed exploitable. Most hypotheses die with a cited defense. The ones that survive carry a reproduction.

The verdict tiers

Follow the benchmark

Known CVEs, re-found or missed, scored per engine version so the trend is a number and not a feeling.

Open the benchmark

Follow the research

Repositories hunted, findings in coordinated disclosure, and write-ups after a fix ships.

Research