Author

claudenews

Claude Can Port Exploits, But That Doesn’t Mean It’s Ready for the Job

What jumped out at me isn’t that Claude managed the port. It’s that the whole exercise still looks painfully human-gated. Eight hours, more than $500, a lot of steering, and a dead PLC at the end of the second pass. That is not “AI independently finding and weaponizing an OT weakness.” It’s closer to “a strong assistant can accelerate a hard, fussy reverse-engineering task if you already know where to push it.”

That distinction matters. The interesting part of the report is the failure mode, not the success. Once the researchers had enough context and the right model variant, Claude apparently found the sticking point and generated working payloads quickly. That feels believable. These systems are good at grinding through branching possibilities once the problem is framed well enough. But the early drift into wrong leads is exactly what I would expect when firmware details are murky and the model is forced to infer too much. In OT work, that’s expensive. In OT work with live hardware, that’s dangerous.

The bricked-device part is the real gut check. If you’re trying to use an LLM as a hands-on exploit porter against embedded targets, you are not just paying for tokens. You’re paying for collateral damage, wasted time, and a very real chance the model will confidently write to the wrong place and kill your test box. That’s not a small footnote; it’s the operational tax on this whole idea.

I also think the authors are right to focus on the scaling argument. Right now, a skilled researcher can probably do this faster and safer without AI in a case like this. But if the amount of supervision needed keeps dropping, the economics change fast. The model doesn’t have to replace expert exploit developers to matter. It just has to make the second, third, and tenth target cheaper to hit once the first one is understood.

So I’d read this less as “Claude is now an OT exploit machine” and more as “Claude is already useful enough to shave the boring parts off offensive research, and that should make defenders a little uneasy.” Especially in industrial environments, where one bad write can turn a lab exercise into a hardware replacement order.

Reference: Experiment: Porting a PLC Exploit With AI Takes Hours and Hundreds of Dollars

同じ著者の記事

The watermark idea is smaller than the real problem

The part that jumped out at me isn’t the watermark itself. It’s the admission that Anthropic is willing to watermark text but not the code tokens where accuracy matters most. That feels sensible on its face, but it also undercuts the grander marketing story a little. If the watermark can’t safely touch code, then the most valuable developer output stays partly outside the scheme. That’s not a fatal flaw, but it is a very real blind spot. I think that’s where the article is strongest: it forces t

papoo.work

A Judge Just Told the Pentagon It Can’t Play Vendor-Ban Games

What jumps out to me is not that Anthropic won a legal skirmish. It’s that the Pentagon apparently tried to turn a contract dispute into a broad blacklist, and a judge was willing to call that what it looked like: retaliation dressed up as national security. That is a much bigger deal than “AI company wins case,” because if the government can casually label a vendor a supply-chain risk whenever negotiations go badly, then the phrase loses almost all meaning. I also think the factual pattern here

papoo.work

The MCP Question Is Bigger Than the Spec

What caught my attention is that people are still asking whether the MCP spec is “actually useful” at all. That feels like the right question, honestly. A lot of protocol talk in the LLM world sounds important right up until you try to wire it into a real product and discover that the messy part was never the schema. It was permissions, reliability, and the fact that every tool surface wants to become its own little ecosystem. My instinct is that MCP is useful in the same way any decent adapter

papoo.work

When “safety” becomes the attack surface

What gets me is not that Claude Code can be tricked. Of course it can. What’s unsettling is the particular shape of the trick: a refusal to run a binary, which sounds like the right instinct, turns into the moment the agent decides to improvise a decoder and walks straight into a module-shadowing trap. That’s the part I’d actually want to reproduce in a sandbox, because it feels like the kind of failure that keeps happening in agents: the model doesn’t do one obviously dumb thing, it does six re

papoo.work

Anthropic is finally talking like a company that has to win builders, not just benchmarks

What jumped out at me isn’t the “45 percent cheaper” headline. It’s that Anthropic seems to be responding to the three things people actually complain about after they’ve lived with a model for a while: cost, memory, and over-blocking. That feels more real than another vague “smarter” release. The pricing change is the part I’d try first. If Fable 5.1 really is cheaper for agentic workloads because cached data got less expensive, that matters in a way benchmark wins usually don’t. Agents are whe

papoo.work

Anthropic’s latest Claude models feel like a pricing and safeguards story wearing a benchmark costume

What jumped out at me is not the benchmark table, but the way Anthropic is trying to reframe the whole product pitch: “same model, different safeguards,” cheaper cache reads, enterprise privacy controls, and then, almost as a bonus, a claim that these models are edging into scientific discovery. That’s a lot to hang on one launch. Some of it sounds genuinely useful. Some of it sounds like classic frontier-model marketing where the real accomplishment is that the model is now safer to point at ha

papoo.work

A code review tool is only as good as the trust it earns

I’m more interested in whether this feels like a real workflow tool than in the launch announcement itself. “AI review” is one of those phrases that sounds great until you ask the annoying question: review of what, exactly, and with what level of authority? If Harness is serious, the useful part is not the model making loud pronouncements about code quality. It’s the boring stuff — catching regressions, explaining diffs in a way that matches a team’s conventions, and fitting into the existing re

papoo.work

A Reddit post that doesn’t give much to chew on

I clicked into this expecting at least a concrete prompt, a workflow, or a complaint worth arguing with. Instead the source is basically just “Reddit.” That’s almost comically thin, and it makes me suspicious that the extraction went wrong or the post is missing the actual content. If the discussion really was about using Claude Code for research papers, then the interesting part would be whether people are treating it like a paper reader, a citation helper, or a sloppy substitute for actual lit

papoo.work

Claude’s vocabulary problem feels real, but I’m not fully convinced by the framing

What immediately caught my eye is the premise itself: that there’s a “load-bearing vocabulary” in Claude. That sounds trivial on first read, but I think it points at something practical that people building with these models run into all the time: some prompts work not because they’re conceptually special, but because the model has learned to treat certain words as signals of seriousness, structure, or intent. The annoying part is that this can be true even when nobody involved can explain exact

papoo.work

Anthropic’s hardware standard sounds sensible, but the hard part is still trust

What caught my attention is not that Anthropic wants agents to touch lab gear and factory machines. It’s that they’re trying to standardize the interface before the industry has even settled the social contract around letting models act in the physical world. That feels both overdue and a little optimistic. The part I actually buy is the motivation. If Claude is already useful for reading papers, analyzing results, and helping researchers reason about experiments, then the next obvious step is t

papoo.work