Did Anthropic just kill appsec engineers?
Code review just got turned into a plugin. 😳
Not only that. This is what security shifting left
I am pretty sure that almost every bit of this post is wrong.
“The ones whose entire value is running SAST and DAST scans, managing tools, and filing tickets? That layer is being automated away. The ones who understand the threat model, think like an attacker, and can direct AI at the right targets? They just got a lot more powerful.”
This is just not accurate. LLMs don’t replace SAST and DAST they augment it. LLMs aren’t automating away PR review or anything else. They are a tool and not even an intelligent one. In fact thanks to LLMs having humans doing review is even more important and the workload isn’t reducing at all. In fact the workload is getting larger. LLMs put out so much rubbish code that it overwhelms people’s ability to properly analyze it.
“Three layers of security review running automatically as Claude writes code: - Pattern matching on every file edit. No AI cost, instant. - A background model review of everything changed each turn. - A deeper agentic review on every commit that reads surrounding code to reduce false positives.”
I can’t imagine what he means by no AI cost. There is a cost and it costs way more than it’s worth. The rest of the text is just bullcrap. Pattern matching and review is just the same approach that has always been used. Instead of LLMs being involved at all it used to be SAST and DAST tools—still should be. At the end of the day LLMs bring pattern matching at scale, but I don't think the frontier models provide value for cost, but I think smaller scale more focused LLMs will remain relevant in the AppSec domain.
I think this kind of post is just the sort of nonsense that makes this LLM bubble so much worse.
Every week, someone in your organization installs an application that hasn't been reviewed, contributing to the growing shadow over your app
This makes me chuckle. I have seen this sort of thing before. I have implemented this before and it is basically a waste of time.
Here is the issue: the review is basically a feel good checkbox that is of negligible value. Whatever risk profile you get out of your review; it is going to be wrong. If you are automating it, then it is performative. If it’s manual, then it’s too expensive.
Furthermore, unless you audit all updates—which you aren’t going to, because it’s to time consuming, the things you “verified” as being okay may well not be after N updates.
The problem is that software is inherently insecure and installing anything is a bit of an unknown risk. Thinking that a review before accepting some application is a security control is just wishful thinking. The review isn’t deep enough and the timeline isn’t long enough.
This is checkbox feel good fake security. You are better off picking a few options and only providing those with no user install capabilities. If you allow for user installs, then you are automatically playing a losing game of whack-a-mole that you can’t control.
The fact is that a security review and assessment actually requires reverse engineering and extensive long term testing. It also needs to be done after every major update. Even with that you aren’t going to know how vulnerable that application is with any real certainty.
To me this is self deluding. You are doing something that gives you a false sense of safety—when in fact it is not much greater safety than what you get from just keeping software up to date.
New results suggest Mythos' cyber threat isn't "a breakthrough specific to one model."
"The new results for GPT-5.5 suggest that, when it comes to cybersecurity risk, Mythos Preview was likely not “a breakthrough specific to one model” but rather “a byproduct of more general improvements in long-horizon autonomy, reasoning, and coding,” AISI writes."
Shocking right? A brute force search is a brute force search. Anthropic isn't unique, and what they are doing isn't some magical Anthropic secret sauce. I guess that is some LLM hype that we can safely dump on the dumpster fire of marketing hype and lies from this LLM bubble.
OpenAI just gave Socket $10M.
Not an investment. A grant. In API credits.
That distinction matters more than people are giving it credit f
This is just such a tiresome narrative. The world is ending but look OpenAI going to save us by investing money they don’t have in technology that isn’t going to change the game now anymore than it has in the past.
Dependency checkers aren’t new. Dependency risk isn’t new. Attackers haven’t been starved for vectors. I feel like the only way you can buy this story is if you believe open source means audited and more secure—which it doesn’t.
The inflection point is here... Nicholas Carlini at Anthropic just had Claude find a zero-day vulnerability live on stage at a conference.
This is the problem logic:
“It ripped through Ghost, an open source project with 50k+ stars on GitHub, zero critical vulns in its entire history. It ripped through it in 90 minutes flat. Then started applying the same logic to beasts like the Linux kernel.”
This isn’t even shocking. This is an issue as old as software itself. The fact that some project hasn’t had a critical vulnerability, even if popular, isn't an indication that it doesn't have any, or make it less likely that someone will find one.
This is a common mistake in logic, and is easily debunked by looking at projects that have been around for decades, and have had vulnerabilities found that have been present for more than ten years.
The fact that GitHub projects have lots of stars doesn't mean that it has been subjected to any real rigor. This is just utterly illogical as a position and doesn't match reality. Software is vulnerable, and always will be. There is always going to be issues lurking where you didn't look, or in forms you didn't think to check for. This is an artifact of software complexity.
"This changes everything. Vuln discovery is now nonstop, automated, and scales like crazy. No fatigue. No clock. Just endless grinding."
This changes nothing. Here is the dirty secret that the LLM bros don't want you to know–vulnerabilities have always been a round the clock risk. Exploit development doesn't stop at 5pm in your timezone. This is not new. LLMs certainly have applications to this field, but their results aren't some indicator of a cosmic shift.
The question isn't can you find a vulnerability–it is can you exploit it reliably? This is something that LLMs aren't great at. They can find vulnerabilities, and they can make exploit chains, but they tend to be weak or not functional.
Remember an LLM is effectively a brute-force search for vulnerabilities. The fact that an LLM can find a vulnerability is not surprising, but they are also not going to be creating novel exploits–since they only can deliver most likely word sequences.
"We are rapidly moving to AI vs AI in cyber security, one hunting, the other racing to patch them live."
No, no we aren't. Do you seriously think this is going to be the case once LLMs are no longer subsidized? LLMs have value right now to experiment with, because they are cheap to use. Once they cost what is profitable–that won't be the case. Once that happens we can be sure that this sort of claim will have zero relevance.
"Remember all those 10-20-30 year old closed source legacy systems that companies are still running... sorry for the sleepless nights with that thought."
This already happens. This is not new. Attackers aren't short of attack surface–they never have been. Why do you think we still have phishing emails today–they work–still. There is a reason why we see the same sorts of attacks again and again–they work and work reliably.
"Use this same tech to scan your code bases, patch ASAP, use AI as a sentry to guard for intrusions and detect early."
Sure. Just subtract the implied claim that LLMs are some sort of security panacea. This is just a wretched post designed to fear monger. If you are an expert, then don't do this. This is how you lose respect.
Update 4/1/2026
Publications from Calif. Contribute to califio/publications development by creating an account on GitHub.
So this is sort of what I mean. People point to things like this and say look at how good LLMs are at exploiting things. Leaving aside the fact that the system exploited has no ASLR; let’s look at the prompts.
Notice the amount of hand holding required to get the LLM to produce this working exploit. People call this an LLM writing an exploit. I guess after a manner of speaking. However, this is just most likely next words based on prompts and memories stored.
The person promoting knew what they were doing and what would and wouldn’t work. The person knew what the expectations were and guided the LLM down the path. This to me is nothing close to the imminent threat some try to claim.
Yes, with hand holding it can exploit things—okay and what is amazing about that. You have an LLM trained on human knowledge, and with hand holding it did something people know how to do and have done. Nothing amazing about that.
The amazing thing might be the fact that it emerged from most likely words picked probabilistically and without understanding. Again as we know as these LLMs get big they gain capabilities up to a point, but they are still pretty feeble on their own.
Cybersecurity firms charge $25K–$150K per pentest.
Someone just open-sourced the entire workflow.
It's called PentAGI.
8,200+ GitHub stars
This is one of those posts that just shows how much of the LLM hype train is at the upper levels of IT. I very much doubt any actual pentesters would say that they can be replaced by an LLM.
I think one of the most shocking quotes is:
“This is the same pattern we see across every industry: Open-source AI doesn't replace experts. It makes the $150K workflow accessible to teams with $0 budgets.”
How does LLM slop make something conducted by an expert accessible for free? Since when were any frontier models free for that use? Since when was cybersecurity so non deterministic?
This post conflates that can is equivalent to should and to done well. I am not sure how anyone in 2026 is still making that mistake. LLMs have found some vulnerabilities and have managed to do some cybersecurity tasks under supervision. However, the general consensus seems to be that they are a far cry from replacing people with real knowledge.
Can someone explain what the number of GitHub stars has to do with quality? OpenClaw has lots of stars and nobody should be using it. If someone can explain this it would be useful? This author seems to conflate it with importance which is also baffling.
Why your Security Architect is actually your best Prompt Engineer.
I’ve spent years writing technical specifications and designing softwar
"Prompt engineering isn't a "soft skill." It's a structural one. If you can write a clear technical specification, you already have the logic required to master Generative AI. We aren't just "talking" to machines; we are programming in prose."
This is just nonsense. This starts with the assumption that an LLM can contribute value to the work of an expert in the domain of cybersecurity. Here is the issue: it almost doesn't matter how fancy your prompts are–the LLM is still going to make mistakes, and you will still have to go through every bit, and make sure it's correct.
At the end of the day your question is: would you rather do the work on the front-end or the back-end? This idea that you are "programming" in prose is just stupid. Programming is a deterministic task. Prompting an LLM is not deterministic. These ideas have nothing to do with each other.
LLMs to an expert is of marginal value. LLMs to non-experts are seemingly valuable. However, the most salient point is that nobody has shown value from LLMs. Businesses claim all this value, and yet they aren't shipping more features or moving into new verticals. The fact is that LLMs have provided zero demonstrable business value. LLMs are built on a cost model that will see them ruined–because nobody will ever pay the unsubsidized rates–given their non-determinist nature leading to an impossible problem of predicting or controlling costs.
Generative "AI" is not changing the world in a lasting way and adjusting your thinking like this–is just making you less valuable in the future, when these products put themselves out of business.
More AI Security Vaporware Based on Nonsense and Magical Thinking
Openclaw immune system v1 built and published for free. 145,000 developers just gave an AI agent root access to their computers. It has no
You can see the code here.
The one good thing I can say is they have code. Okay good news out of the way. The bad news is this is vaporware. There isn’t even a product or a viable thing you can run to improve security. The author claims experience–long-term experience in AI but the stuff here shows no evidence. This seems to be written assuming that LLMs are somehow closer to human consciousness than they are.
The utter non-seriousness of this is evidenced by bits like this:
Inhale: I am. Exhale: still here.
How is this sort of thing supposed to be remotely useful to something that works at a semantic level? The very language used throughout is the way you might talk to a kid. This again just proves how off base this is. LLMs are statistical chatbots that act at semantic level. They don't feel or think. They generate the most likely words. An LLM can't breath and it can't derive comfort from anything.
This LLM craze is just bringing out all the rubbish ideas. The amusing fact is that this idea fits with OpenClaw–they are both terrible; just for different reasons.
