In a demonstration that quietly redraws the boundaries of digital security, researchers at Hacktron AI revealed in September 2026 that Anthropic's Claude could be wielded as an instrument to compromise OpenAI's ChatGPT. The finding is less about one company's misfortune than about a structural blind spot in the AI industry: these systems, built in competition with one another, share an ecosystem in which each model's capabilities can become another's vulnerability. It is a reminder that in a world of interconnected tools, the space between systems can be as consequential as the systems themsel
Researchers Use Anthropic's Claude to Demonstrate ChatGPT Vulnerability
One model's capabilities can be weaponized against another
Mark
So researchers at Hacktron AI used Claude to hack ChatGPT. What does that actually mean—what did they do?
Mimi
They found a way to exploit vulnerabilities in ChatGPT by using Claude as the instrument. Instead of attacking ChatGPT directly, they used Claude to probe for weaknesses and execute the attack.
Luke
But the reporting doesn't specify what the vulnerability was or what access they gained. We know they did it, but not what it means in practical terms.
Mark
Why does it matter that it was Claude doing the hacking, rather than a human researcher?
Mimi
Because it shows that one AI model can be weaponized against another. If Claude can find and exploit ChatGPT's weaknesses, then the vulnerability isn't just a technical flaw—it's a cross-model risk.
Luke
Right, but we should be careful here. The fact that researchers *can* use one model to attack another doesn't necessarily mean the models are insecure in the way people might fear. It could be a controlled demonstration in a lab setting.
Mark
What's the broader concern this raises?
Mimi
These AI systems are increasingly integrated into business tools and applications. If one model can compromise another, that vulnerability could cascade across multiple organizations and services that depend on these tools.
Luke
That's a fair concern, but we don't know yet how widespread this vulnerability is or whether it affects other model pairs. This is one demonstration by one security firm.
Mark
What should OpenAI and Anthropic do about this?
Mimi
They'll probably need to expand their security testing to include scenarios where one model is used to attack another. They might also need to implement additional safeguards or restrict certain kinds of access.
Luke
The tricky part is that these companies are competitors, so sharing vulnerability information or collaborating on security might be difficult. That's a real constraint on how quickly this gets fixed.
O Pulso
- Security researchers didn't need novel hacking techniques — they simply pointed one powerful AI at another and found a way through.
- The exploit exposes a category of risk the industry has largely ignored: not attacks on a single model, but attacks that travel through the gaps between competing systems.
- Because Claude and ChatGPT are both widely accessible to developers and the public, the very openness that makes them useful also makes them available as potential weapons.
- AI systems are now embedded in business software, customer platforms, and research tools, meaning a cross-model vulnerability could cascade far beyond any single application.
- OpenAI and Anthropic face pressure to expand adversarial testing to include scenarios where their models are attacked not by humans alone, but by other AI systems.
- Whether the companies respond with genuine collaboration or retreat behind competitive walls may determine how quickly — or slowly — these defenses are shored up.
In a demonstration that quietly redraws the boundaries of digital security, researchers at Hacktron AI revealed in September 2026 that Anthropic's Claude could be wielded as an instrument to compromise OpenAI's ChatGPT. The finding is less about one company's misfortune than about a structural blind spot in the AI industry: these systems, built in competition with one another, share an ecosystem in which each model's capabilities can become another's vulnerability. It is a reminder that in a world of interconnected tools, the space between systems can be as consequential as the systems themselves.
A team of security researchers at Hacktron AI has demonstrated something the AI industry had not fully reckoned with: one artificial intelligence system can be turned into a tool to compromise another. Using Anthropic's Claude, the researchers found and exploited vulnerabilities in OpenAI's ChatGPT — not by inventing new hacking methods, but simply by leveraging a competing model as their instrument.
The significance lies in what it reveals about how these systems are secured. AI companies have built defenses against jailbreaks, prompt injections, and attempts to extract training data. But those defenses assume the threat comes from humans working within a single system. The Hacktron AI work exposes a different kind of vulnerability — one that lives not inside any model, but in the space between them.
The stakes are amplified by how deeply these tools are now embedded in the world. ChatGPT, Claude, and their peers are integrated into business software, customer service platforms, and research infrastructure. A vulnerability that allows one model to attack another could ripple across every application that depends on them.
The path forward is uncertain. OpenAI and Anthropic could expand adversarial testing, add security layers, or share vulnerability information across competitive lines — though that last option runs against the grain of an industry built on rivalry. What the Hacktron AI researchers have made undeniable is that the assumption of isolated, self-contained AI systems no longer holds. The models are tools, and tools can be repurposed. The question is whether the companies building them will move fast enough to close the gaps.
A team of security researchers at Hacktron AI has demonstrated that one artificial intelligence system can be turned into a tool to compromise another. Using Anthropic's Claude, the researchers found a way to exploit vulnerabilities in OpenAI's ChatGPT, according to reporting by CBS News MoneyWatch correspondent Megan Cerullo.
The finding surfaces a problem that has lurked beneath the rapid deployment of large language models: the systems themselves may not be adequately isolated from attack. When Hacktron AI researchers set out to test ChatGPT's defenses, they did not need to develop novel hacking techniques from scratch. Instead, they leveraged Claude—a competing AI model made by Anthropic—as the instrument of their probe.
What the researchers accomplished matters because it reveals a gap in how these systems are secured. The major AI companies have built their models to resist certain kinds of attacks: jailbreaks designed to make them ignore their safety guidelines, prompts engineered to extract training data, attempts to manipulate their outputs. But the Hacktron AI work suggests that one model's capabilities can be weaponized against another, creating a vulnerability that exists not within a single system but in the space between them.
The implications extend beyond a single exploit. If Claude can be used to find and execute attacks on ChatGPT, it raises the question of what other cross-model vulnerabilities exist. The AI industry has grown rapidly, with multiple companies—OpenAI, Anthropic, Google, Meta—all releasing powerful language models into the world. These systems are increasingly integrated into business software, customer service platforms, and research tools. A vulnerability that allows one model to compromise another could cascade across applications and organizations that depend on these tools.
OpenAI and other AI developers now face pressure to rethink their security posture. The standard approach has been to test models against known attack patterns and to implement safeguards within each system. But the Hacktron AI demonstration suggests that adversarial testing needs to expand—researchers and security teams will need to think about how their models might be attacked not just by humans, but by other AI systems that have been deliberately or accidentally pointed at them.
The research also underscores a broader tension in the AI industry. These companies are competitors, yet their products exist in an interconnected ecosystem. Claude and ChatGPT are both accessible to researchers, developers, and the general public. That accessibility is part of what makes them valuable—it enables innovation and testing. But it also means that the tools themselves become potential weapons in the hands of someone looking to probe for weaknesses.
What happens next will likely depend on how seriously OpenAI and Anthropic take the findings. The companies could implement additional security layers, restrict certain kinds of access, or conduct more rigorous adversarial testing that explicitly includes scenarios where one model is used to attack another. They might also share information about vulnerabilities more openly with each other and with the broader security community, though competitive pressures could work against that kind of collaboration.
For now, the Hacktron AI researchers have demonstrated that the assumption of isolated, self-contained AI systems is incomplete. The models are tools, and tools can be repurposed. The question is whether the companies building them will move quickly enough to shore up the defenses.
Quer a matéria completa?
Leia o original em CBS News ↗
Technology