AI processor as a symbol of autonomous, AI-driven cyber attacks and AI-assisted penetration tests
Technical ArticlesT-46124 Jul 20266 min

Agentic AI in cyber security: when attackers go autonomous – and why defenders still need to hold the reins

The cyber security industry has been debating the potential of artificial intelligence for years – as a tool for attackers as well as defenders. A recent incident has given this debate a new and very concrete dimension. At the same time, at DFIR Austria and Probatio Cyber Defence we are currently gaining intensive hands-on experience with AI-assisted penetration tests – and the results paint a nuanced, in parts surprising picture. The two developments belong together, because they show where agentic AI actually stands on the maturity curve today, both in offence and in defence.

The JadePuffer case: the first fully AI-driven ransomware attack

With JadePuffer, security researchers at Sysdig have documented an incident considered one of the first of its kind: a ransomware attack carried out from start to finish by an autonomous LLM agent – without a human operator having to direct individual steps.

The attack at a glance

  1. Initial access: the agent exploited CVE-2025-3248, an unauthenticated remote code execution vulnerability in Langflow, a widely used open-source framework for LLM applications. The flaw was patched back in April 2025 and listed by CISA as actively exploited shortly afterwards – yet many instances remain reachable on the internet, unhardened and unpatched.
  2. Reconnaissance & credential access: after gaining code execution, the AI independently searched Langflow’s PostgreSQL database, collected host information, combed through environment variables and sensitive files, and enumerated a MinIO object store.
  3. Adaptive error handling: what stands out is how the agent dealt with unexpected situations. When an interface unexpectedly returned XML instead of the expected JSON, the AI adapted its parsing logic on its own in the next attempt. In another sequence, a failed login attempt was turned into a working exploit within just 31 seconds.
  4. Persistence: on the Langflow host, the agent installed a cron job that contacted the attacker’s infrastructure every 30 minutes.
  5. Lateral movement & privilege escalation: from there, the agent moved to a production MySQL server running Alibaba Nacos and exploited CVE-2021-29441, an authentication bypass vulnerability, to create rogue admin accounts.
  6. Impact: finally, the AI encrypted 1,342 Nacos configuration objects using MySQL’s AES_ENCRYPT(), deleted the original tables and left behind a purpose-built extortion table with a ransom demand, a Bitcoin address and a Proton Mail contact.

Traces of AI control

The details that point to AI control are just as interesting: the generated code contained extensive natural-language comments on the agent’s operational logic, and according to Sysdig the AES-256 encryption claimed in the ransom note is more likely the weaker AES-128-ECB – a detail that suggests an AI-generated implementation reproducing “plausible” statements from its training data without verifying them technically. The Bitcoin address in the ransom note is also a sample address widely used in documentation – probably another artefact of the model’s training data.

Sysdig sums up the significance of the case aptly: we are at the beginning of the era of “agentic threat actors” – autonomous AI agents that replace the technical skill that used to be required for a successful attack. At the same time, the characteristic patterns of AI-generated payloads also open up new detection opportunities for defenders.

The other side of the coin: our hands-on experience with AI-assisted pentests

While JadePuffer shows what agentic AI on the attacker side is already capable of under favourable conditions, at DFIR Austria and Probatio Cyber Defence we are gathering our own experience on the defensive side. We are currently running our fourth AI-assisted penetration test for a client – in both an external and an internal (client-side) scenario. The picture that emerges is considerably more mixed than the JadePuffer case would suggest.

External: fast, structured, convincing

In the external pentest, the AI works impressively efficiently. It identifies attack surfaces quickly, structures its approach in a traceable way and delivers reliable, consistent results. This is where the potential seen in the JadePuffer case shows: clearly scoped, well-documented target systems with known vulnerability patterns are an environment in which agentic AI moves with great confidence.

Internal: inconsistent and resource-hungry

The internal, client-side scenario is a different story. The findings are much less consistent and in some cases do not really fit together. For six hours of testing we measured consumption of around 200,000 tokens – simply too much for practical, economical use at this stage.

Our interim conclusion after four runs

AI-assisted tools already provide very good support for clearly scoped basic tasks – such as BloodHound analysis or Active Directory checks, where the job is to work systematically through known patterns. But as soon as no clear direction is given, the AI gets lost in complex internal networks surprisingly quickly, wandering into all kinds of dead ends. Exactly the focused, fault-tolerant adaptability that made the JadePuffer agent so effective as an attacker is still missing from current AI pentesting tools in this form – at least in more complex, less standardised environments.

What does this mean for organisations?

Putting both experiences side by side yields an important insight for corporate security strategy: agentic AI is no longer a theoretical vision of the future but already usable in production in certain contexts – on both sides of the front line.

  • On the attacker side, this means: internet-facing, poorly hardened systems with known, publicly documented vulnerabilities are now a realistic target for fully automated, AI-driven attacks. Patch management, attack surface reduction and securing systems that hold credentials and API keys are more important than ever.
  • On the defender side, this means: AI-assisted pentests can already speed up testing meaningfully for clearly structured, standardised tasks. In complex internal networks with organically grown, heterogeneous structures, however, AI does not (yet) replace the experience and direction of qualified pentesters – it complements them.

For us as a provider of penetration testing, managed SOC services and digital forensics, this means in concrete terms: AI belongs in the toolbox of our analysts and pentesters, but – as things stand – not yet in the driving seat of a complex internal assessment. Among attackers, who are actively testing this boundary right now, the situation is exactly the reverse.

Conclusion and outlook

The JadePuffer case and our own test series are two sides of the same stage of development: agentic AI can already execute complete attack chains autonomously when the conditions are right – exposed systems, clear error patterns, documented vulnerabilities. In more complex, less predictable environments, such as typical internal corporate networks, the same technology still runs into clear limits.

We will continue to follow this development in our ongoing test series and share our findings regularly. For organisations that want their own security posture reviewed – whether through classic or AI-assisted penetration tests, managed SOC services or digital forensics – we are an experienced partner. Get in touch.