Using Claude to reduce the attack surface of code

Oct 21, 2026•dnw•View Original

TL;DR Highlight

Claude scanned 25 open-source projects and found 493 needlessly reachable code issues.

Who Should Read

Backend and security engineers with complex server configs who can't tell what code runs before authentication.

Core Mechanics

Shrink reachable code, not just bugs
- Finding and fixing individual vulnerabilities scales poorly because it says nothing about the remaining ones.
- Reducing reachable code lowers the severity of unknown vulnerabilities too.
493 issues across 25 projects
- The targets were 25 open-source projects in C, C++, Java, PHP and Python, over five million lines.
- Claude inventoried over 17,000 config options, of which 3,842 were security relevant.
Three lenses for finding them
- Code reachable in the default config but unused by default functionality.
- Option pairs that look fine alone but cause problems together.
- Complex, sensitive code that runs before authentication, without victim interaction, or before a protocol state transition.
Config combos silently disabled security
- 162 security-sensitive combos were accepted with no warning or docs, and 94 of them silently turned off a security control.
- In gated projects, 93 unnecessary pre-authentication surfaces were found in default configs.
- Hostname deny rules failed open when reverse DNS was off, and this pattern appeared in four unrelated projects.
Even OpenSSH had some
- In 10.4, about 4,500 lines of experimental ML-DSA-44 code were reachable by unauthenticated clients in the default config.
- That was about 16% of default sshd's pre-authentication code.
- The key was parsed before the allow check, and 10.5 moved the check ahead of parsing.
Packaged as a reusable Claude skill
- It runs three passes, each starting with an empty context plus a short summary from the previous pass.
- Each finding is confirmed by a dynamic trace, and default-config gates must match the project's test suite and a black-box regression suite.
- Gates that would change non-default configs are set aside for a human.

Community Reactions

Many complained AI refuses security work
- One commenter said Claude and GPT refused to audit their own codebase, so GLM or Kimi finished the job.
- Another said Claude and Codex balked at PHP CVE investigation and at writing auth-related unit tests, so they switched to GLM 5.3.
- Another said Claude drops to less capable models on security topics even with Cyber Verification Program membership.
Doubts about result quality
- One commenter cited a kernel developer's talk claiming most earlier AI-found kernel vulnerabilities were non-issues or already fixed.
- They wondered whether AI labs rush to announce results without understanding the underlying problems.
Someone tried the same approach
- One commenter used the prompt "reduce attack surface, move features into feature-flag Rust crates, build a minimal version" on Vix IDE.
- The codebase was split into crates and Cargo features so an auditor can review one capability at a time.
Counterpoint on AI-written code
- One commenter said Anthropic discusses finding bugs in others' code but not the vulnerabilities AI creates in new code.

How to Apply

List your pre-auth entry points
- Today, open your codebase in Claude Code and have it list the parsers and handlers that run before auth or session setup.
- Mark whether each is reachable in the default config and whether it is actually used.
Audit security option combinations
- List security-related options and test them in pairs for silently disabled controls.
- Check whether listeners or login paths remain when a feature is disabled.
- Reject bad combos with an error at config load.
Move checks before parsing
- Put allow-list checks ahead of deserialization and signature verification.
- Compare the type the parser actually uses, not what the client claims.
Verify gates with regression tests
- After adding a gate, run the existing test suite to confirm default behavior is unchanged.
- Route changes that affect non-default configs to a human for approval.

Try It

copy & run

# Paste into Claude Code, Claude or ChatGPT (simplified version of the article's skill)

You are a security reviewer. Find candidates for attack surface reduction in the codebase below. Work in three passes and end each pass with a 3-line summary for the next.

Pass 1: Code reachable in the default config but unused by intended functionality
Pass 2: Complex code that runs before authentication, without user interaction, or before a protocol state transition
Pass 3: Option combinations that are fine alone but silently turn off or weaken a security control together

In each pass, first build an inventory of entry points. For each finding, tabulate file/function, reachability path, and a proposed gate (reorder checks, state-order guard, reject combo at config load). Mark proposals that change default behavior as 'needs human decision'.

Target code: {{codebase or directory}}

Next step: apply one proposed gate and run {{test command}} to confirm default behavior is unchanged.

Terminology

Attack surfaceThe total of code and input paths an attacker can touch. Like a house, more doors and windows mean more ways in.

Pre-authenticationThe stage before login or authentication completes. Like a building lobby anyone can enter, so code here is risky.

Fail openWhen a security control fails, it lets traffic through instead of blocking. Like a broken door lock that stays open.

GateA check placed before code runs. It is like putting a door in front of the code so it runs only after auth or explicit opt-in.

ML-DSA-44A post-quantum signature algorithm. In OpenSSH 10.4 it was experimental code.

Claude skillA reusable bundle of instructions that teaches Claude a repeatable workflow.

Related Resources

Using Claude to reduce the attack surface of code

Oct 21, 2026•dnw•View Original

TL;DR Highlight

Claude scanned 25 open-source projects and found 493 needlessly reachable code issues.

Who Should Read

Backend and security engineers with complex server configs who can't tell what code runs before authentication.

Core Mechanics

Shrink reachable code, not just bugs
- Finding and fixing individual vulnerabilities scales poorly because it says nothing about the remaining ones.
- Reducing reachable code lowers the severity of unknown vulnerabilities too.
493 issues across 25 projects
- The targets were 25 open-source projects in C, C++, Java, PHP and Python, over five million lines.
- Claude inventoried over 17,000 config options, of which 3,842 were security relevant.
Three lenses for finding them
- Code reachable in the default config but unused by default functionality.
- Option pairs that look fine alone but cause problems together.
- Complex, sensitive code that runs before authentication, without victim interaction, or before a protocol state transition.
Config combos silently disabled security
- 162 security-sensitive combos were accepted with no warning or docs, and 94 of them silently turned off a security control.
- In gated projects, 93 unnecessary pre-authentication surfaces were found in default configs.
- Hostname deny rules failed open when reverse DNS was off, and this pattern appeared in four unrelated projects.
Even OpenSSH had some
- In 10.4, about 4,500 lines of experimental ML-DSA-44 code were reachable by unauthenticated clients in the default config.
- That was about 16% of default sshd's pre-authentication code.
- The key was parsed before the allow check, and 10.5 moved the check ahead of parsing.
Packaged as a reusable Claude skill
- It runs three passes, each starting with an empty context plus a short summary from the previous pass.
- Each finding is confirmed by a dynamic trace, and default-config gates must match the project's test suite and a black-box regression suite.
- Gates that would change non-default configs are set aside for a human.

Community Reactions

Many complained AI refuses security work
- One commenter said Claude and GPT refused to audit their own codebase, so GLM or Kimi finished the job.
- Another said Claude and Codex balked at PHP CVE investigation and at writing auth-related unit tests, so they switched to GLM 5.3.
- Another said Claude drops to less capable models on security topics even with Cyber Verification Program membership.
Doubts about result quality
- One commenter cited a kernel developer's talk claiming most earlier AI-found kernel vulnerabilities were non-issues or already fixed.
- They wondered whether AI labs rush to announce results without understanding the underlying problems.
Someone tried the same approach
- One commenter used the prompt "reduce attack surface, move features into feature-flag Rust crates, build a minimal version" on Vix IDE.
- The codebase was split into crates and Cargo features so an auditor can review one capability at a time.
Counterpoint on AI-written code
- One commenter said Anthropic discusses finding bugs in others' code but not the vulnerabilities AI creates in new code.

How to Apply

List your pre-auth entry points
- Today, open your codebase in Claude Code and have it list the parsers and handlers that run before auth or session setup.
- Mark whether each is reachable in the default config and whether it is actually used.
Audit security option combinations
- List security-related options and test them in pairs for silently disabled controls.
- Check whether listeners or login paths remain when a feature is disabled.
- Reject bad combos with an error at config load.
Move checks before parsing
- Put allow-list checks ahead of deserialization and signature verification.
- Compare the type the parser actually uses, not what the client claims.
Verify gates with regression tests
- After adding a gate, run the existing test suite to confirm default behavior is unchanged.
- Route changes that affect non-default configs to a human for approval.

Try It

copy & run

# Paste into Claude Code, Claude or ChatGPT (simplified version of the article's skill)

You are a security reviewer. Find candidates for attack surface reduction in the codebase below. Work in three passes and end each pass with a 3-line summary for the next.

Pass 1: Code reachable in the default config but unused by intended functionality
Pass 2: Complex code that runs before authentication, without user interaction, or before a protocol state transition
Pass 3: Option combinations that are fine alone but silently turn off or weaken a security control together

In each pass, first build an inventory of entry points. For each finding, tabulate file/function, reachability path, and a proposed gate (reorder checks, state-order guard, reject combo at config load). Mark proposals that change default behavior as 'needs human decision'.

Target code: {{codebase or directory}}

Next step: apply one proposed gate and run {{test command}} to confirm default behavior is unchanged.

Terminology

Attack surfaceThe total of code and input paths an attacker can touch. Like a house, more doors and windows mean more ways in.

Pre-authenticationThe stage before login or authentication completes. Like a building lobby anyone can enter, so code here is risky.

Fail openWhen a security control fails, it lets traffic through instead of blocking. Like a broken door lock that stays open.

GateA check placed before code runs. It is like putting a door in front of the code so it runs only after auth or explicit opt-in.

ML-DSA-44A post-quantum signature algorithm. In OpenSSH 10.4 it was experimental code.

Claude skillA reusable bundle of instructions that teaches Claude a repeatable workflow.

Using Claude to reduce the attack surface of code

TL;DR Highlight

Who Should Read

Core Mechanics

Community Reactions

How to Apply

Try It

Terminology

Related Papers

Related Resources

Using Claude to reduce the attack surface of code

TL;DR Highlight

Who Should Read

Core Mechanics

Community Reactions

How to Apply

Try It

Terminology

Related Papers

Related Resources