BẢN CHÉP ÂM THANH 🔗
WEBVTT
00:00:00.239 --> 00:00:11.439
So what if an AI model could spend days working independently on a complex software engineering problem, identify vulnerabilities that human developers might miss, and then write the code needed to fix them?
00:00:11.599 --> 00:00:21.039
Now imagine that that same AI can also help identify weaknesses in the very systems organizations rely on to protect their businesses, their customers, and their data.
00:00:21.280 --> 00:00:24.640
Well, Google has introduced a model designed to do exactly that.
00:00:24.800 --> 00:00:34.079
It's called Gemini 4 Carbon, and it's making some impressive claims, particularly when it comes to cybersecurity, software engineering, and autonomous AI agents.
00:00:34.320 --> 00:00:44.799
However, only a couple of weeks ago I covered Google's Gemini 4 Argon, which was already being positioned as a powerful frontier model for software engineering, enterprise work, and cybersecurity.
00:00:45.119 --> 00:00:48.159
So why has Google introduced another model so quickly?
00:00:48.399 --> 00:00:52.479
What exactly is Gemini 4 carbon and what makes it different from Argon?
00:00:52.880 --> 00:01:06.959
And perhaps most importantly, could this be a significant step forward towards AI systems that don't simply help us write software, but increasingly take responsibility for finding problems, solving them, and delivering complex engineering work independently.
00:01:07.200 --> 00:01:17.359
And in this episode, I'll look at some of these genuinely impressive performance figures, and there are some important questions around security, control, and what it means for the future of technology teams.
00:01:17.599 --> 00:01:20.719
Welcome to the Inspiring Tech Leaders with me, Dave Roberts.
00:01:21.040 --> 00:01:32.560
So Google announced Gemini 4 Carbon on the 28th of October 2026, positioning it as the latest frontier model and the strongest model yet for cybersecurity and complex long-running tasks.
00:01:32.799 --> 00:01:42.079
The announcement comes just two weeks after the introduction of Gemini 4 Argon, demonstrating how quickly Google is moving to develop increasingly capable AI systems.
00:01:42.239 --> 00:01:46.640
But carbon isn't simply another incremental update to an existing chatbot.
00:01:46.879 --> 00:01:58.239
Google is positioning it as a model designed to tackle challenging engineering problems that require sustained reasoning, multiple stages of work, and the ability to recover when things don't go according to plan.
00:01:58.560 --> 00:02:01.280
Think about a major software development project.
00:02:01.439 --> 00:02:08.879
You might have millions of lines of code, complicated dependencies, legacy systems, and bugs that are extremely difficult to identify.
00:02:09.120 --> 00:02:22.560
Traditionally, resolving those problems requires experienced developers spending considerable time investigating the underlying causes, testing different solutions, and making sure that fixing one problem doesn't create another one somewhere else.
00:02:22.800 --> 00:02:35.360
Google says carbon can work on these types of challenges for extended periods, with its internal teams reporting it that it has continued working for days, navigating obstacles and attempting to resolve complex engineering problems.
00:02:35.599 --> 00:02:39.120
That is a significant change in how we should think about AI.
00:02:39.439 --> 00:02:43.919
We're moving beyond asking a model to generate a piece of code or answer a question.
00:02:44.080 --> 00:02:52.479
We're increasingly talking about AI that can take responsibility for a much larger task, work through a series of challenges, and deliver a completed outcome.
00:02:52.639 --> 00:02:59.199
And that's where the distinction between a conventional AI assistant and an autonomous AI agent become particularly important.
00:02:59.680 --> 00:03:03.919
So let's go back to Gemini 4 Argonne, which I covered on the show a few weeks ago.
00:03:04.240 --> 00:03:11.840
Argon was introduced as Google's frontier model for complex workflows across software engineering, enterprise knowledge work, and cybersecurity.
00:03:12.000 --> 00:03:18.319
It was designed to identify software vulnerabilities, investigate potential weaknesses, and generate solutions.
00:03:18.560 --> 00:03:28.639
It also demonstrated impressive capabilities in understanding attack services and producing proof-of-concept exploits to establish whether vulnerabilities could actually be exploited.
00:03:28.879 --> 00:03:35.840
In other words, Argon was already a serious proposition for organizations working in cybersecurity and software development.
00:03:36.080 --> 00:03:38.719
So where does Carbon fit into that picture?
00:03:39.039 --> 00:03:52.319
The easiest way to understand the difference is to think of Argonne as an already highly capable frontier model, whereas Carbon represents Google's next step towards more sustained, reliable, and security-focused autonomous engineering.
00:03:52.560 --> 00:04:03.840
Carbon builds on Argon's capabilities but puts considerably more emphasis on long-running engineering tasks, generating secure code, identifying vulnerabilities and producing fixes.
00:04:04.080 --> 00:04:10.639
Google says Carbon is more than twice as successful as Argon at generating secure code in its internal benchmarks.
00:04:10.800 --> 00:04:18.319
It has also improved its ability to recognise when it doesn't have enough information rather than simply producing an answer that sounds convincing.
00:04:18.560 --> 00:04:25.759
Google reports that carbon produces approximately half as many unsupported claims as Argon in its internal testing.
00:04:26.000 --> 00:04:36.879
And that's important because one of the biggest challenges with AI-generated code isn't simply whether the code works, it's whether the code is reliable, secure, and actually does what the developer intended.
00:04:37.120 --> 00:04:40.560
Carbon is also designed to work more
Load full transcript
collaboratively with people.
00:04:40.720 --> 00:04:51.040
When faced with ambiguity or conflicting information, it's more likely to identify the problem, investigate for available evidence, and ask for clarification rather than making assumptions.
00:04:51.360 --> 00:04:59.120
So this isn't just about making AI faster or more intelligent, it's about making it more dependable when working on complicated tasks.
00:04:59.360 --> 00:05:08.639
Now, when it comes to Gemini 4 carbon, cybersecurity is one of the main areas Google has focused on, and the company has published some impressive benchmark results.
00:05:08.959 --> 00:05:18.480
On the Artificial Analysis Intelligence Index version 5, Carbon achieved a score of 59, placing it at the top of the benchmark at the time of its announcement.
00:05:18.639 --> 00:05:26.639
That put it ahead of several competing Frontier models, including Claude Opus 5.5 and OpenAI's GPT-6 Astra.
00:05:27.040 --> 00:05:30.959
But the cybersecurity results are arguably even more significant.
00:05:31.199 --> 00:05:38.720
On Secbench Pro, which measures a model's ability to identify real-world software vulnerabilities, Carbon achieved 88.2%.
00:05:39.199 --> 00:05:46.399
On ExploitGym, a benchmark focused on establishing whether vulnerabilities can actually be exploited, it achieved 54.8%.
00:05:47.360 --> 00:05:54.240
And on CWE Bench, which measures the ability to address weaknesses through software patches, carbon achieved 71%.
00:05:55.120 --> 00:06:00.800
What makes these results interesting is that they cover different stages of cybersecurity process.
00:06:01.040 --> 00:06:10.800
First, identifying a vulnerability, then determining whether that vulnerability can genuinely be exploited, and finally producing a software fix to address the underlying problem.
00:06:11.040 --> 00:06:16.399
That's a much broader capability than simply scanning code and highlighting something that might be wrong.
00:06:16.639 --> 00:06:23.040
It suggests that AI models are becoming increasingly capable of supporting the complete vulnerability management process.
00:06:23.279 --> 00:06:26.720
However, it's important to put these figures into context.
00:06:26.959 --> 00:06:33.040
Benchmark performance doesn't automatically translate into equivalent results in a live enterprise environment.
00:06:33.279 --> 00:06:41.759
Real-world systems are considerably more complicated, and organizations still need independent testing, human oversight and proper security controls.
00:06:42.000 --> 00:06:44.639
Nevertheless, the direction of travel is clear.
00:06:44.959 --> 00:06:49.120
AI is becoming a much more capable participant in software security.
00:06:49.439 --> 00:06:53.040
For technology leaders, the implications could be significant.
00:06:53.279 --> 00:06:59.199
One of the biggest challenges facing organizations today is the growing complexity of their technology environments.
00:06:59.439 --> 00:07:07.040
Businesses are running increasingly complicated applications, cloud platforms, interconnected services, and legacy systems.
00:07:07.279 --> 00:07:11.279
At the same time, the number of potential vulnerabilities continues to grow.
00:07:11.519 --> 00:07:19.519
Security teams are under pressure to identify and prioritize risks, while development teams are expected to deliver new capabilities more quickly.
00:07:19.759 --> 00:07:27.040
There simply aren't enough highly experienced security specialists and software engineers available to deal with every challenge manually.
00:07:27.199 --> 00:07:30.319
This is where models such as carbon could make a real difference.
00:07:30.560 --> 00:07:42.800
Imagine being able to deploy an AI agent to investigate a complex code base, identify potential vulnerabilities, establish which ones represent the greatest risks, and recommend or generate appropriate fixes.
00:07:43.040 --> 00:07:54.480
Instead of developers spending hours investigating individual problems, they could potentially focus their attention on reviewing the findings, validating the solutions, and making decisions about what should happen next.
00:07:54.639 --> 00:08:02.480
That could improve productivity, reduce the time to address vulnerabilities, and allow security teams to concentrate on the risks that matter the most.
00:08:02.720 --> 00:08:05.199
But there's another important consideration.
00:08:05.519 --> 00:08:14.160
If AI can help defenders find and fix vulnerabilities, it can also make it easier for attackers to discover weaknesses and develop more sophisticated attacks.
00:08:14.399 --> 00:08:22.319
The same underlying capabilities that make carbon attractive to cybersecurity teams could potentially be misused if they fall into the wrong hands.
00:08:22.560 --> 00:08:26.160
This is why Google's approach to deployment is particularly important.
00:08:26.720 --> 00:08:30.800
Google isn't making Gemini 4 carbon immediately available to everyone.
00:08:30.959 --> 00:08:40.720
The initial rollout is restricted to a selected group of cybersecurity defenders with access gradually expanding into trusted partners through the Google Fairwind program.
00:08:41.120 --> 00:08:51.840
The company says it's taking a deliberate approach to deployment, continuing to evaluate the model and strengthen its safeguards before making it more widely available to developers, businesses and consumers.
00:08:52.159 --> 00:08:56.159
And there's some interesting developments in the security measures themselves.
00:08:56.399 --> 00:09:00.080
One of the areas Google has concentrated on is prompt injection.
00:09:00.320 --> 00:09:15.039
For anyone unfamiliar with that term, prompt injection is where malicious instructions are introduced into information an AI system is processing, potentially attempting to manipulate the model into ignoring its original instructions or performing actions it shouldn't.
00:09:15.360 --> 00:09:22.960
This becomes particularly concerning when an AI agent has
access to files, applications, systems, or sensitive business information.
00:09:23.279 --> 00:09:32.639
Google says carbon achieved a 0% attack success rate across the evaluated attempts on the Grace One indirect prompt injection benchmark.
00:09:32.879 --> 00:09:38.559
That's an improvement on Argon, which already achieved a very low attack success rate of 0.7%.
00:09:39.600 --> 00:09:48.559
Google has also reported improvements in reducing attempts to take shortcuts, staying within the boundaries of an assigned task and avoiding unsupported claims.
00:09:48.879 --> 00:09:57.679
These are important developments because the more responsibility we give to AI agents, the more important it becomes to ensure they operate safely and predictably.
00:09:57.919 --> 00:10:06.480
However, a 0% result on a particular benchmark doesn't mean a model is completely immune to prompt injection or other security attacks.
00:10:06.720 --> 00:10:13.360
There is no substitute for proper security architecture, access controls, monitoring, and human accountability.
00:10:13.679 --> 00:10:22.960
Technology leaders need to think carefully about what permissions AI agents receive, what systems they can access, and what actions they are allowed to perform independently.
00:10:23.279 --> 00:10:25.840
So what about availability and pricing?
00:10:26.080 --> 00:10:32.879
At launch, Gemini 4 Carbon is being made available to a limited group of cybersecurity defenders through the Fair Win program.
00:10:33.279 --> 00:10:41.440
Google intends to expand access to additional trusted partners before making it available more broadly through its developer, business and consumer products.
00:10:41.679 --> 00:10:44.000
The initial pricing is interesting though.
00:10:44.240 --> 00:10:56.240
Google has announced an introductory price of 2 US dollars per million input tokens and 10 US dollars per million output tokens, matching the introductory pricing of Gemini 4 Argon.
00:10:56.720 --> 00:11:06.240
After the introductory period, the published pricing is expected to increase to$4 per million input tokens and$20 per million output tokens.
00:11:06.559 --> 00:11:16.879
For organizations considering using these models at scale, pricing is obviously an important consideration, but I would argue that the bigger question is the overall return on investment.
00:11:17.120 --> 00:11:28.159
If an AI model can significantly reduce the time required to investigate vulnerabilities, resolve software defects, and improve the quality of development work, the potential savings could be substantial.
00:11:28.320 --> 00:11:37.200
Of course, those benefits need to be measured against implementation costs, infrastructure, governance, and the need for experienced people to validate the results.
00:11:37.519 --> 00:11:47.200
The real opportunity isn't necessarily replacing developers or cybersecurity professionals, it's enabling them to achieve considerably more with the resources they already have.
00:11:48.000 --> 00:11:55.440
So stepping back, I think the introduction of Gemini 4 carbon tells us something important about where the AI industry is heading.
00:11:55.679 --> 00:12:04.799
We're seeing a move away from models that simply respond to instructions towards models that can take increasingly complex objectives and work towards completing them.
00:12:04.960 --> 00:12:13.440
The ability to sustain work over extended periods, understand complicated software environments, and recover from problems is particularly important.
00:12:13.919 --> 00:12:16.080
And this isn't happening in isolation.
00:12:16.320 --> 00:12:24.720
Google's announcement of carbon also coincided with the introduction of Genie4, a separate development focused on interactive world simulation.
00:12:25.200 --> 00:12:27.440
So what should we take away from all of this?
00:12:27.679 --> 00:12:40.320
Gemini 4 Carbon isn't simply another version of Google's AI chatbot, it's a model designed to tackle complex engineering challenges, work for extended periods, and deliver significant improvements in cybersecurity.
00:12:40.480 --> 00:12:51.840
Compared with Gemini 4 Argon, Carbon represents a clear step forward in secure cogeneration, long-running autonomous tasks, reliability, and resistance to prompt injection.
00:12:52.080 --> 00:13:00.480
The benchmark results are impressive, particularly its ability to identify vulnerabilities, validate potential exploits, and generate software patches.
00:13:00.720 --> 00:13:09.759
But I think the most important development is the shift towards AI that can undertake substantial pieces of work rather than simply assisting with individual tasks.
00:13:10.080 --> 00:13:15.840
For technology leaders, that creates enormous opportunities, but it also introduces new responsibilities.
00:13:16.080 --> 00:13:25.279
We need to understand how these models work, where they can add value, how we govern their actions, and most importantly, how we maintain appropriate human oversight.
00:13:25.440 --> 00:13:36.000
Because the question is no longer simply whether AI can write code, it's whether we can trust AI to investigate, develop, test, and secure the systems that our businesses depend upon.
00:13:36.240 --> 00:13:38.559
And that's a much bigger conversation.
00:13:38.799 --> 00:13:40.080
Well, that's all for today.
00:13:40.159 --> 00:13:42.559
Thanks for tuning in to the Inspiring Tech Leaders Podcast.
00:13:42.720 --> 00:13:48.080
If you found this episode useful, make sure you subscribe, turn on notifications, and share the episode with your network.
00:13:48.159 --> 00:13:53.039
You can find more insights, show notes, and resources at www.inspiringtechleaders.com.
00:13:53.200 --> 00:13:58.240
Head over to the social media channels you can find Inspiring Tech Leaders on X, Instagram, Inspo, and TikTok.
00:13:58.399 --> 00:14:03.759
Thanks for listening, and until next time, stay curious, stay connected, and keep pushing the boundaries of what's possible in tech.