Fixing one vulnerability says nothing about the ones that remain. Can a model instead shrink the amount of code an attacker can reach, so that unknown vulnerabilities become harder to exploit? This post, published on 20 October 2026 by Chris Rohlf of Anthropic's Frontier Red Team, previews "initial research into automated attack surface analysis and reduction" (introduction). The attack surface of a program, as the post uses the term, is the code an attacker can reach; reducing it means moving code that a deployment does not need behind a gate, such as authentication, an opt-in setting or a later protocol state. The post is the version of record. It has no companion paper or released dataset, and the OpenSSH release notes it links are the outside record of its case study. The post is about 2,500 words long.
The three analyses
The post defines three ways to find unnecessary reachable code ("Methods for reducing codebases' attack surface"):
- Default configuration. Code reachable in the shipped default configuration that no default functionality uses.
- Configuration interaction. Pairs of options that look reasonable alone but, together, silently disable a security control or leave a feature's behavior undefined while its code stays reachable.
- Early reachability. Security-sensitive, complex code that runs before a security control such as authentication, without victim interaction, or before a protocol state transition.
The team applied these to 25 open-source projects, "over five million lines of code," in languages "including C, C++, Java, PHP, and Python." Claude built an inventory of over 17,000 configuration options, of which 3,842 related to a security-relevant control or feature. Projects were grouped by intended exposure: Public (serving anyone, 10 projects), Gated (serving authenticated users, 11) and Private (serving the operator's own environment, 4).
What Claude found
The post reports 493 issues in total ("Claude's findings"). It does not say who confirmed them or by what standard; the only confirmation step it describes belongs to the reusable skill in its closing section (below), which confirms each finding "with a dynamic trace of the program," and the post does not say the 493 were found with that skill. No independent review or false-positive rate is reported.
- Gated projects. All 11 had unintentionally reachable code, though never "a significant percentage of the total lines of code." In this cohort Claude found 93 unnecessary pre-authentication surfaces in the default configuration, 37 places where code ran or state was recorded before authentication, and 61 configuration combinations that silently weakened an administrator's pre-authentication control. The post says none of these "would have been immediately exploitable on their own."
- Configuration interactions, all projects. 162 security-sensitive combinations were accepted without error, warning or documentation of how they interact. In 94 the combination silently turned off a security control; the other 68 were inconsistencies in the same mechanisms, such as a listener that stayed reachable when the configuration said it was disabled.
- A recurring pattern. In four unrelated projects, an access rule written with a hostname could never match when reverse DNS lookups were off (the setting most projects used), so a deny list "would fail open." Some projects documented the dependency on lookups; none stated the fail-open consequence.
- Conflicting options. In one project, an option meant to defer a client-certificate check instead disabled it, so a server configured for trusted clients accepted anyone. The combination had been present "for 21 years."
- Namespace-dependent settings. In another project, a client-certificate requirement was silently ignored on some endpoints, depending on which form of the setting name the administrator used.
What the chart shows. The post's figure, "Issues found per project by three attack surface analyses," plots counts per project and analysis without printing values. Read from its gridlines (this note's reading), the totals by analysis are 225 default-configuration, 162 configuration-interaction and 106 early-reachability issues, summing to 493. The configuration total matches the 162 in the text. By cohort, the readings are about 185 issues for Public projects, 239 for Gated and 69 for Private. The Gated cohort's chart counts (107, 82 and 50 by analysis) are larger than the text's 93, 61 and 37, which describe the pre-authentication subsets; the post does not reconcile the two.
The OpenSSH case
Measured as a share of the code, the largest reduction was in OpenSSH ("Case study: OpenSSH"). The account, in order:
- OpenSSH 10.4. Claude found roughly 4,500 lines of new, experimental code for ML-DSA-44, a post-quantum signature algorithm, reachable by an unauthenticated client in the default configuration. ML-DSA was not enabled by default, but the daemon parsed whatever key type a client sent before checking whether that type was allowed. The post puts the 4,500 lines at "about 16% of OpenSSH's own pre-authentication code in a default sshd."
- OpenSSH 10.5. After Anthropic's report and patches, the release moved the check before parsing. Claude then found two gaps: the check read the algorithm name the client advertised, not the type inside the key itself, and certificate authority signatures were still verified for any supported algorithm before the configuration was consulted.
- OpenSSH 10.6. After several iterations with the maintainers, the checks moved inside the parser. The post says the fix "removed roughly a fifth of the OpenSSH code that an unauthenticated client previously could have reached in a default build," almost all of it the ML-DSA code.
The OpenSSH release notes confirm the sequence (this note's check). Release 10.4 (6 July 2026) added the composite ML-DSA-44 and Ed25519 scheme as experimental and "not enabled by default." Release 10.5 (11 August 2026) moved the key-type check before parsing, "Suggested by Christopher Paul Rohlf of Anthropic." Release 10.6 (6 October 2026) checks key and CA signature types during parsing, crediting Rohlf "in collaboration with Claude and Anthropic Research." The same 10.6 notes also list, under new features, "enable hybrid post-quantum ssh-mldsa44-ed25519 signature algorithm." Portable release 10.6p2 (5 November 2026, after the post) lists among its bugfixes "enable the post-quantum hybrid signature scheme ssh-mldsa44-ed25519 as an allowed CA signature algorithm." The post does not discuss how enabling the algorithm bears on its reachability figures.
The 10.5 and 10.6 notes open with the same paragraph on AI-generated reports in general: many "are determined not to have security impact when considered in the context of a realistic threat model," and the team welcomes them "especially when combined with human triage."
The reusable skill
The post says the three analyses have been combined into "one reusable Claude skill," which Anthropic is working to integrate into its OSS Scanner ("Conclusion"). The skill runs three passes, each with a fresh context window apart from a short handoff from the previous pass. In each pass Claude inventories entry points and confirms each finding with a dynamic trace. For default-configuration findings, it implements a gate and requires the project's test suite, plus a black-box regression suite, to give identical results to the baseline. A gate that would change a non-default configuration is "set aside for a human to decide." For the other two analyses it proposes changes such as reordering a check or rejecting a combination when the configuration loads. The skill itself is not published.
Instructions and choices described
The post publishes no prompts. The three analysis definitions above are, in effect, the task descriptions given to the model ("how well Claude could take these three descriptions of attack surface and discover real issues"), and the post quotes them only in summary.
Limitations
- No grading. The 493 issues are Claude's findings. The post does not say how they were confirmed; the skill it describes later confirms findings with Claude's own dynamic traces. The post gives no human review, no false-positive count and no maintainer acceptance rate, except in the OpenSSH case and in the general statement that patches were sent.
- Unnamed projects and model. Apart from OpenSSH, the 25 projects are not named, and the post does not say which Claude model was used.
- No comparison. There is no baseline from static analysis tools or human auditors, and no cost figure, though the post says Claude can work "economically."
- Severity. The post itself says none of the Gated issues was immediately exploitable alone; the benefit claimed is a reduction in exposure, which is not measured against any later exploit.
- Chart and text. The per-cohort counts in the text and the chart measure different subsets, as noted above.
Research questions
- What share of the 493 issues would maintainers accept? Submitting the findings and recording acceptance, rejection and "won't fix" outcomes would estimate the skill's precision.
- Does reducing reachable code lower later exploitation? Tracking vulnerabilities found later in code that had been gated, against code left reachable, would test the post's central claim.
- How does the method compare with existing tools? Running static reachability analysis or configuration linters on the same 25 projects would show what the model adds.
- Does the test-suite equivalence check catch behavior changes? Measuring how often gated builds that pass the project's tests still break real deployments would test the skill's safety check.
- How common is the fail-open hostname pattern? A survey of projects that accept hostname-based access rules would show whether four of 25 is representative.
- What does the 10.6 enabling of ML-DSA do to the reachable-code figures? Re-measuring pre-authentication reachable code in default builds of OpenSSH 10.4, 10.5 and 10.6 would settle the "fifth" figure across releases.
How it connects
- The post links Anthropic's earlier capability reports: Assessing Claude Mythos Preview’s cybersecurity capabilities, Partnering with Mozilla to improve Firefox's security (March 2026) and Measuring LLMs’ ability to develop exploits. It presents attack surface reduction as the defensive step beyond finding and exploiting individual bugs.
- It names An opt-in vulnerability-finding service for open-source software (October 2026) as the service the skill is being integrated into.
- Discovering and exploiting vulnerabilities in critical, real-world software with Claude (October 2026), with the same author among its contributors, reports the offensive side of the same period's work; the connection is this note's.
References
- Rohlf, C. (2026, October 20). Using Claude to reduce the attack surface of code. Anthropic Frontier Red Team. https://www.anthropic.com/research/attack-surface-reduction record
- OpenSSH. Release notes, OpenSSH 10.4 (2026-07-06), 10.5 (2026-08-11), 10.6 (2026-10-06) and 10.6p2 (2026-11-05). https://www.openssh.org/releasenotes.html record