· via dev.to (home feed)
SSRF flaw in Continue's docs crawler affects versions 2.0.0 and 2.1.0
Continue versions 2.0.0 and 2.1.0 include a documentation crawler vulnerable to server-side request forgery, letting a crafted URL trigger requests to unintended internal destinations.

What happened
A server-side request forgery (SSRF) vulnerability has been disclosed in the documentation crawler shipped with Continue, the AI coding assistant. According to a security write-up on dev.to, the flaw is tracked as CVE-2026-133695 and affects versions 2.0.0 and 2.1.0 of the tool.
The vulnerable code sits in DocsCrawler at core/indexing/docs/crawlers/DocsCrawler.ts. The public description, as relayed by the dev.to post, states that a URL fed to the crawler can trigger requests to destinations the user never asked for, meaning the indexing component can be steered toward hosts that an outside party should not be able to reach through it.
Severity depends on which scoring model you consult. The dev.to article reports a CVSS v3 base score of 5.0 against a CVSS v4 score of 2.3, a gap it attributes to how the newer model weighs attack prerequisites and user interaction.
How the flaw becomes a real problem
A documentation indexer exists to fetch remote sites and follow links so it can build a searchable corpus. That fetcher runs either on a developer's machine or on a shared service, and as the dev.to author points out, the process can often reach localhost, internal package registries, cloud metadata endpoints or company wikis from there.
When the crawler follows a URL supplied by an attacker without checking where the request actually resolves, the request is made from a position of trust. The attacker needs no inbound access to the internal service, because the crawler opens the outbound connection on their behalf. Depending on the tool's behavior, the response might be indexed, written to logs or leaked through an error message; even without seeing response bodies, connection failures and timing differences can hint at whether a target exists, the author notes.
The upstream repository is read-only
Complicating remediation, the dev.to post reports that Continue's repository is now read-only. That removes the default assumption that an upstream patch will arrive on any predictable schedule. The article suggests fixes may instead come from a maintained fork, a downstream package, or an organization patching its own copy, and cautions that the state of the public repository is no guarantee that security fixes will be provided.
Mitigations worth applying
The write-up lays out a defense-in-depth approach for teams that keep using the crawler:
- Disable documentation indexing entirely if it is not needed.
- Otherwise, restrict the crawler to an allowlist of public domains, and resolve hostnames before connecting.
- Block loopback, link-local, private, multicast and cloud metadata address ranges.
- Re-check every redirect and every DNS resolution, because a single hostname can point at several addresses.
- Run the crawler in a network namespace with no route to internal management networks.
- Strip cloud instance credentials from the process, or require a metadata proxy that enforces workload identity.
- Treat HTTPS as no proof that a destination is safe.
For monitoring, the author recommends logging requested URLs, resolved addresses, redirect chains and final statuses, while avoiding sensitive page bodies, and alerting whenever a request is blocked for a private address. Repeated attempts from a single user or project merit investigation.
On verification, the post advises testing against public documentation, a loopback URL, a private RFC1918 address, a redirect that lands on a private address, and a hostname resolving to both public and private records. Every non-public case should fail before any connection is opened, and the documentation index should still build normally once the policy is enabled.
Why it matters
Two things lift this above a routine CVE listing. First, developer tooling tends to run close to valuable targets: internal registries, wikis and cloud metadata services are precisely what SSRF is designed to probe, and a coding assistant's docs crawler is a natural vehicle for it. Second, the read-only upstream repository means anyone running 2.0.0 or 2.1.0 cannot simply wait for the next release; they need to inventory where the tool runs, apply network controls themselves, or move to a maintained fork. The split CVSS scores are also a useful reminder that severity is model-dependent: a team triaging on v4 numbers might deprioritize what v3 rates as a medium-severity issue.
- #security
- #ssrf
- #vulnerability
- #developer-tools
- #ai-coding-assistant