Introducing cve.now: A Free Feed of Security Vulnerabilities for Agents

Four Major Security Incidents in a Single Day

October 21st was a rough day for anyone running software in production.

First, we learned about PHP2Shell: two bugs in core PHP that make it possible for attackers to run commands on your server with a single POST request. Amazon reported an exploit attempt four days prior to disclosure, which means attackers have a head start.

Then there was the LiteLLM supply chain attack. Attackers managed to publish a malicious version of LiteLLM to both PyPI and Docker, where it was downloaded by thousands of users. This version, 1.107.0, steals cloud and AI credentials from anyone unfortunate enough to run it.

There was also a Node.js security update (versions 26.12.1, 24.22.1, and 22.23.4), which addressed 10 security vulnerabilities. One of these stands out as particularly severe: a Brotli decoder bug that allows for remote code execution.

Finally, the Railway hack. A Railway disclosure published on Wednesday, October 21 reported unauthorized access to Railway's internal systems, including customer environment variables, API keys, and DB credentials. Much of the industry spent the next 24 hours rotating keys.

If this is the new normal, how are software teams expected to keep up?

The answer might seem obvious. Many of these issues show up first as a CVE or security advisory. If you subscribe to the sources that publish this information, you can find out fast and have your agent check whether you're affected, pushing patches when necessary.

Unfortunately, this isn't as simple as it sounds.

CVE Data Sources Are Challenged

CVE data sources have major problems that make it hard for software teams to use them effectively. There are three main ones today.

First, data is fragmented.

• For the main registry and identifier system, you need the CVE List.
• For a list of vulnerabilities that have actually been exploited, you need CISA KEV.
• EPSS contains the likelihood that a CVE will actually be exploited (if it hasn't been already).
• OSV contains the list of affected packages and versions.

If you don't have all four, it's hard to evaluate risk and make prioritization decisions.

Second, there's no “trigger.” In general, agents only act when you ask them to, which means a critical CVE can sit for hours before a team member notices and directs their agent to investigate. Running your agent on a loop helps, but this approach is costly, and you still lose time between intervals.

Third, the sheer number of vulnerabilities is overwhelming. More than 48,000 CVEs were published last year - over 130 a day. In 2026, the number is certain to be much higher. Most of these CVEs don't apply to you, but without effective filters, it's hard to identify the ones that do.

Introducing cve.now

cve.now is a free feed of CVEs and security advisories that addresses these problems.

1. It does the joining for you. Every CVE is a single entry with four sources merged in, along with additional sources that we add every day. It includes a compact, plain-English summary.
2. It tells you when to look. cve.now exposes a public RSS feed in multiple formats, including JSON, so you can subscribe to vulnerabilities as they are announced. Agents can access the feed via MCP. Humans can pipe alerts to Slack or Microsoft Teams.
3. It helps you cut through the noise. cve.now allows you to construct filtered feeds based on specific technologies, programming languages, package managers, etc. It even provides a utility for generating the relevant filter set based on your application's lockfile.

Most importantly, cve.now is free for anyone to use. No login, signup, or API key required.

How It Works

Every 5 minutes, cve.now pulls new and updated records from the CVE List, CISA KEV, EPSS, and OSV. It merges these into one entry per CVE by cross-checking identifiers - never by fuzzy-matching. This means a CVE and its associated GitHub advisory will never show up twice, and two different bugs will never collapse into a single entry.

Merging is valuable not just for convenience, but also for determining which vulnerabilities actually matter. Take CVE-2025-61624, a Fortinet bug from last week. Its severity score is 5.4, a “medium” - but CISA says attackers are actively exploiting it. An agent looking only at the CVE record would tell you to ignore it. On cve.now, it appears in the critical feed, making it hard to miss.

The critical feed is deterministic, not AI-driven. cve.now considers an entry to be “critical” if any of the following are true:

1. It appears in CISA's KEV catalog. This means someone is actively exploiting it.
2. It has a CVSS score of 9.0 or higher and affects a widely used technology. We decided that a high score alone is not enough because many 9.0+ CVEs appear in software that almost nobody runs. That's why we add a “widely used in” heuristic, which is based on a curated list of about 30 technologies such as Next.js, Node.js, Kubernetes, NGINX, and PostgreSQL.
3. It has a CVSS score of 7.0 or higher and a high likelihood of being exploited soon. This means an EPSS score of at least 0.1, which puts it in roughly the top 5% of all CVEs.

It's important that the critical feed is deterministic, not AI-driven, because error and inaccuracy are unacceptable in this context. It also helps with cost: by subscribing to the critical feed, you can avoid paying an agent to triage tens or hundreds of CVEs every day.

After merging the data, cve.now uses AI to write a short, plain-English summary of each entry. The summary is labeled as AI-written, and the original description is always included right below it. Facts like severity, exploitation status, and affected versions always come from the source data, never the model.

Everything is served as RSS, Atom, and JSON feeds, filterable by technology or package ecosystem (e.g., just PHP, or just npm). There's also a JSON API and an MCP server.

How You Can Use It

cve.now is easy to use.

You can add the MCP server to your coding agent by pasting this prompt into Claude Code, Codex, Cursor, or any other agent that supports MCP:

Add the cve.now MCP server to your MCP configuration: name "cve-now", remote Streamable HTTP server at https://cve.now/mcp, no authentication. Then confirm the tools get_vulnerability, list_feed and check_stack are available, and run check_stack on this project's lockfile to list any known vulnerabilities in my dependencies.

Afterwards, your agent can look up any CVE and check your dependencies against recent vulnerabilities on its own.

If you use Slack, you can install Slack's RSS app, then subscribe to a channel:

/feed subscribe https://cve.now/feed?filter=critical

New critical CVEs will show up in your Slack channel as they appear.

Limitations

cve.now is powerful but not without limitations. Here are the key ones today:

1. History: cve.now has about the last 30 days of the CVE List. There are many thousands of CVEs that were published before this cutoff date. Be aware: if cve.now finds nothing for your dependencies, that doesn't mean you're safe from older vulnerabilities. We may consider backfilling the entire dataset soon.
2. Version matching: Version matching works for package ecosystems that publish version ranges, like npm, PyPI, Go, Maven, and crates.io. It doesn't cover Nix: a flake.lock pins source trees rather than packages, so nixpkgs package CVEs aren't checked.
3. Summaries: AI summaries are written for new entries as they arrive. Some older entries may still show the original description.
4. SBOM uploads and OS packages: They're not included today. An SBOM (software bill of materials) is a file, usually in CycloneDX or SPDX format, that lists every component in a build. OS packages are the system packages your servers and containers install with apt, apk, or yum. cve.now checks the dependencies in your lockfile; checking SBOMs and OS packages needs Linux distribution advisory data, which we haven't added yet.

Last, and this one is very important: cve.now does not include information or entries for incidents that never received a CVE. Many security incidents, including the recent Railway hack, do not qualify as CVEs and will never receive this assignment. These incidents will not show up in cve.now today - or in any vulnerability feed, for that matter. We may consider incorporating them in the future, but for now, you should not treat cve.now as your exclusive source of security notices.

Why We Built This

Security is top of mind for everyone right now, including Sazabi. Whenever a major security incident occurs, our customers ask the same question: “Are we affected?” Sazabi aims to help them answer this question as quickly as possible.

Just as important are the vulnerabilities that our customers may not be aware of. Again, Sazabi has a responsibility to notify our customers about these and help them implement mitigations as quickly as possible.

cve.now is how we deliver on this promise. By consuming signals from cve.now, Sazabi is able to automatically check if customers are affected by new security vulnerabilities in real time and notify them immediately. The cve.now feed and MCP server are enabled by default for every Sazabi customer and project.

We recognized that this data source would be useful to everyone, not just Sazabi customers, so we made it free and publicly available.

If you have questions about cve.now, please refer to the website. If you would like to file a bug or feature request, you can do so by opening an issue on the GitHub repository or by tweeting at me: @shcallaway.

Sherwood Callaway
CEO & Founder