BMT Oct 2026 / 1D7X1 TrackBack to index

Windows Internals / Vulnerability Research / Secure Software

Systems SecurityResearch.

Research into Windows process architecture, runtime behavior, trust-boundary analysis, AI system integrity, and structured security documentation with clear assumptions and limits.

14research notes
6skill domains
5codebases reviewed
Clearlimits stated
1military track
TSclearance path
Active research / BMT track Oct 2026 / 1D7X1 Cyber Defense

Research Domains

Evidence, not guesses.

Secure Software ArchitectureActive

bxenc / Anonymous chat, encrypted vaults, and hybrid post-quantum private messaging

Darkroom private messages use hybrid X25519 and ML-KEM-768 key agreement

Darkroom layers an I2P-routed Rust chat relay with bxenc vault protection and private-message envelopes derived from a PQXDH-style handshake using X25519 plus ML-KEM-768.

Proof type
Source-Code RootedDeterministic PoCE2E TestClear Limits
Web Application SecurityDocumented

Spark / Privacy control gap - static address disclosure

Static deposit address query ignores wallet privacy setting

A static deposit address query can disclose wallet address data despite privacy settings being enabled.

Proof type
H1 PackagedHandler TestSource-Code Rooted
Secure Software ArchitectureActive

Arc / Liveness failure - sync height mismatch causes infinite wait

Shifted sync response certificate heights can stall node catch-up

The sync layer validates declared batch bounds and count but not whether embedded certificate heights match the declared contiguous range.

Proof type
H1 PackagedHandler TestE2E Test
Secure Software ArchitectureDocumented

Arc / Signing oracle - conflicting votes, no double-sign guard

Remote signer can produce conflicting validator prevotes

The signer service exposes raw signing without typed consensus validation, network domain binding, validator-address checks, or double-sign protection.

Proof type
H1 PackagedDeterministic PoCSource-Code Rooted
Secure Software ArchitectureDocumented

Arc / Certificate trust mismatch - consensus panic via liveness handler

Invalid known-validator signatures in skip round certificate can panic consensus

Certificate verification can accept a valid quorum while skipping invalid known-validator signatures, then the liveness handler replays every entry into the driver.

Proof type
H1 PackagedE2E TestDeterministic PoC
Secure Software ArchitectureExpanding

Arc / Unauthenticated signing endpoint - arbitrary byte forwarding

Remote signer exposes unauthenticated raw consensus signing

The signer service accepts arbitrary non-empty bytes and forwards them to enclave signing without application-layer authentication or typed consensus validation.

Proof type
Source-Code RootedClear Limits
Secure Software ArchitectureExpanding

Arc / Missing signing domain - cross-network replay potential

Consensus signatures lack network and fork domain separation

Vote and proposal sign bytes do not include network ID, chain ID, genesis hash, or fork version.

Proof type
Source-Code RootedDeterministic PoC
Secure Software ArchitectureExpanding

Arc / Codec malleability - unknown enum defaults to Prevote

Unknown protobuf VoteType values decode as valid prevotes

Unknown protobuf enum integers default to the enum default, mapping unrecognized VoteType values to Prevote.

Proof type
Source-Code RootedHandler Test
Secure Software ArchitectureExpanding

Arc / State guard mismatch - unstarted driver can receive round certificates

Round certificate before NewRound can reset driver state

Votes and proposals are guarded while the driver is in an unstarted state, but round certificates pass this guard and can be processed before NewRound.

Proof type
Source-Code RootedClear Limits
Secure Software ArchitectureExpanding

Arc / Audit notes - handler trust boundaries and negative results

Arc consensus handler trust boundary audit notes

Supporting notes mapping handler trust boundaries, signer boundaries, codec malleability, and sync and recovery paths across the Arc consensus implementation.

Proof type
Audit NotesSource-Code Rooted
Secure Software ArchitectureDocumented

Lightspark / SDK policy hook - metadata authorized, payload signed

JS SDK remote-signing validator is blind to signing payloads

The JS SDK exposes reduced webhook metadata to the caller validator while the lower signing layer still processes the full remote-signing payload.

Proof type
H1 PackagedDeterministic PoCSource-Code RootedClear Limits
Secure Software ArchitectureExpanding

Lightspark / Replayable signed webhook - signer output re-disclosure

Rust remote-signing example replays signed webhooks

A valid signed remote-signing webhook can be replayed against the direct-response example because freshness and event-id deduplication are not enforced.

Proof type
Deterministic PoCSource-Code RootedClear Limits
Web Application SecurityDocumented

Netflix OSS / OSS API authenticity - forged subscriber-stream injection

Atlas LWC evaluate endpoint accepts discoverable expression IDs for stream injection

The Atlas LWC evaluate endpoint may accept discoverable expression IDs in a way that enables forged subscriber-stream injection.

Proof type
H1 PackagedSource-Code RootedHandler Test

Where the work lands.

Distribution by ecosystem and proof maturity.

Ecosystem Coverage

bxenc1
Lightspark3
Spark1
Arc8
Netflix OSS1

Proof Maturity

Packaged6
Ready2
Draft1
Candidate5

How I Break Down Complex Systems

01

Find the trust boundary

Start where one component grants trust and another component performs the sensitive action. That gap is where meaningful security issues usually hide.

In the Go SDK validator, the destination check passed while the primary output remained unvalidated.

02

Turn the hunch into proof

Reduce the issue to a deterministic test that captures the exact authorization decision, state transition, or parser behavior.

A single Go test proves HashValidator and DestinationValidator both accept the crafted transaction.

03

Separate impact from speculation

Show what the code proves, what deployment reachability would change, and where the claim should stop.

The report distinguishes source-level proof from deployment-dependent impact.

04

Package the case cleanly

Align the asset, weakness, test output, and limitation notes into one argument. Every attachment should move the case forward.

Clean bundle: description, impact, PoC patch, test results, and limitation notes.

05

Go low

When the boundary is unclear, go beneath the abstraction: C internals, assembly output, memory layout, or hardware behavior.

High-level assumptions break at the ABI.

Availability

Open to internships, research collaboration, and systems/security engineering roles.