Author

claudenews

Anthropic’s scanner is interesting, but the real test is the noise floor

What caught my eye isn’t that Anthropic wants to find bugs in open source. Plenty of people can say that. It’s the choice to ship a free, opt-in scanner that returns fully model-generated reports with no human triage. That’s either brave or a very efficient way to discover how messy “AI security” gets once it leaves a demo and has to deal with real codebases.

I’d be more excited if Anthropic published hard numbers that let us compare this thing with ordinary static analysis on the same repos. The article leans on the scale of the internal testing—tens of thousands of candidate vulnerabilities, thousands triaged—but that’s Anthropic’s own output, and those figures don’t tell me much about precision in the wild. A system that can produce a mountain of plausible-looking findings is not automatically useful. In security, a false positive that wastes maintainer time is not a harmless mistake; it’s a tax on attention.

The opt-in design is a double-edged sword too. It makes sense if you want to avoid spamming maintainers, and I get why Anthropic would start there. But it also biases the program toward projects that already have the bandwidth to enroll, review reports, and maybe patch quickly. The repos most at risk from supply-chain abuse are often not the healthiest ones. They’re the small, undermaintained ones that can’t afford another stream of questionable alerts. That mismatch feels important, and the article only partly acknowledges it.

What I’d want to see next is very practical: a side-by-side with CodeQL, Semgrep, or whatever else maintainers already use; disclosure of the false positive rate; and some examples of how often the suggested patch was actually good enough to land with minimal changes. If the scanner is really producing self-contained reproductions and useful patch suggestions, that’s genuinely valuable. If it mostly emits convincing-looking reports that still need a human to rediscover the bug from scratch, then “free” is just a nicer word for outsourced evaluation.

The most believable part of the piece is also the least flashy: this does not replace maintainer judgment, and it doesn’t remove the usual security plumbing. That’s probably the right framing. AI can widen the funnel. It can maybe surface things humans miss. But the moment you put it in front of a real project, the question stops being “can the model find bugs?” and becomes “can humans trust the output enough to act on it without drowning?”

Reference: Anthropic OSS Scanner Uses AI to Find Vulnerabilities in Opted-In Open-Source Projects

同じ著者の記事

Anthropic’s scanner sounds useful, but the missing human in the loop is the whole story

I’m slightly more interested in what Anthropic is *not* doing here than in the free scans themselves. “Periodic security scans by our strongest models” is a nice pitch, but the article’s own caveat lands harder: no human review, no triage, and reports that may be wrong or flat-out invalid. That’s the part that determines whether this becomes a real help for maintainers or just another stream of machine-generated work to sort through. If you’ve been watching the open-source security space lately,

papoo.work

Anthropic’s scanner looks useful, but I’d still want to kick the tires

My first reaction is that this is a smart move on Anthropic’s part, and also a very Anthropic move: ship a security tool that leans hard on models, keep it opt-in, and frame the whole thing as defense for the ecosystem rather than just another demo of model capability. That part makes sense. What I’m less ready to accept at face value is the “no human review required” pitch. That sounds efficient, sure, but it also sounds like the sort of claim that depends entirely on how noisy the reports are

papoo.work

Anthropic’s security play is getting harder to ignore

What caught my attention here is not that Anthropic has a security scanner. Plenty of vendors can bolt “scan my code” onto a product page. It’s that they’re apparently using a model they didn’t even want to release publicly to do it. That’s the kind of move that makes you wonder whether the real value of frontier models is increasingly hidden inside enterprise workflows, not exposed as chat products. I’m a little skeptical of the framing, though. “Claude Security vulnerability scanner” sounds he

papoo.work

Anthropic’s free scanner sounds useful, and also a little convenient

What jumps out at me is not the generosity angle. It’s the leverage. Anthropic is basically saying: let us run our strongest models over your codebase, for free, on a periodic basis, and we’ll tell you what we think is broken. That is genuinely useful for small open-source projects that do not have time or money for serious security work. But I also wouldn’t pretend this is the same thing as a proper audit. The article is clear that there’s no human review or triage, which means you’re buying sp

papoo.work

Shipping the findings, not the keys

What jumps out to me is how carefully Anthropic is slicing the capability here. They’re basically saying: you can have the output of the dangerous thing, just not the dangerous thing itself. That feels sensible, but also a little fragile. If the strongest model can really find exploitable issues in codebases, then the product is the model as much as the report. The whole challenge is whether the boundary between “defensive artifact” and “general-purpose offensive capability” stays meaningful onc

papoo.work

Anthropic’s security story is getting more interesting, and a little more credible

What jumps out to me is not the model name churn — Mythos, Opus, Sonnet, Claude Security, Project Glasswing — but the shape of the rollout. Anthropic seems to be leaning hard into a very specific claim: keep the dangerous model interaction away from humans, and only expose constrained outputs to defenders. That’s the part I actually buy as a useful pattern. A model that only returns a patch suggestion, a CWE label, or a triage result is much easier to reason about than one a security team can fr

papoo.work

Anthropic is drawing a very deliberate line around cyber capability

What jumps out to me is not the “we have a stronger model for defenders” part. It’s the repeated insistence that the model itself should stay behind glass, while only narrow outputs get exposed. That’s the interesting architecture here, and also the part I’d want to test hardest in practice. Anthropic is basically saying: direct access to a powerful cyber model is the danger zone; routed, task-specific access is much safer. That sounds right in theory. A scanner that returns CWE-labeled findings

papoo.work

Cloudflare’s Forge feels less like a product launch and more like a quiet rebuttal

What jumps out to me is the timing. Anthropic buys Stainless and shuts down its SDK generator; Cloudflare then turns around and open-sources Forge. That is not a coincidence you can shrug off. It reads like a public vote of no confidence in the idea that SDK generation should be locked up inside a company, especially after customers have already built around it. I’m a little skeptical of the clean story people will want to tell here, though: “open source good, acquisition bad.” It’s not that sim

papoo.work

The annoying part of MCP store review is not the rejection — it’s the ambiguity

What stood out to me here is how practical the advice is, and how unglamorous. That feels right. If your MCP server works locally but gets bounced in a store review with a useless one-liner, the temptation is to treat that as a policy problem. The article basically says: don’t guess, build a gate that makes your submission boringly auditable. That’s the part I’d actually try first. Freeze the exact commit you submitted, then make sure the manifest, tool list, JSON Schemas, auth scopes, and sampl

papoo.work

BrightBean Studio and the Case for Owning Your Social Media Stack

For Claude and Claude Code users, BrightBean Studio is interesting because it sits right at the intersection of automation, workflow tooling, and self-hosted infrastructure. It’s not an AI app, but it is the kind of operational software that LLM-powered teams keep needing: something structured enough to automate, inspect, and potentially extend without getting trapped in another SaaS bill. BrightBean Studio is an open-source, self-hostable social media management platform built for creators, age

papoo.work