AI-Assisted Hacking: Separating Marketing Hype from Technical Reality

The intersection of Large Language Models (LLMs) and cybersecurity has sparked a polarized debate. On one side, headlines suggest a new era of "vibe hacking" and solo actors breaching government infrastructures using AI; on the other side, seasoned engineers argue that the hype far outweighs the actual technical capability of these models.

Understanding where AI actually adds value in the exploit chain—and where it fails—is critical for security professionals trying to determine if their defensive posture needs a fundamental shift.

The Gap Between Benchmarks and Real-World Code

A recurring theme in the analysis of AI-driven attacks is the disparity between performance in controlled environments and performance against "messy" real-world codebases.

For example, Anthropic's SCONE benchmark (December 2025) demonstrated a significant leap in capability. AI agents successfully exploited 19 out of 34 post-cutoff smart contracts—a 55.8% success rate. This is a staggering increase from the roughly 2% success rate observed just twelve months prior. These agents were able to simulate the theft of $4.6M in funds at an average API cost of only $1.22 per contract.

However, these successes are concentrated in narrow, well-scoped vulnerability classes, specifically Solidity smart contracts. When shifted to general-purpose, complex software, the results are less impressive. Daniel Stenberg recently tested a gated Anthropic frontier model against the curl codebase. Despite the hype, the model produced five "confirmed" findings, only one of which was triaged as a low-severity CVE. Stenberg noted that the surrounding hype was "primarily marketing," suggesting that while LLMs can find trivial bugs, they struggle with the deep architectural logic required to breach hardened, professional-grade software.

The Exploit Supply Chain

Beyond the technical ability to find a bug, the actual execution of a breach involves a complex supply chain designed to minimize liability. The process of exfiltration and monetization is rarely handled by a single entity.

1. The Discovery: An actor may use an LLM to identify a vulnerability. Because this creates a digital paper trail with the AI provider, the discoverer may choose not to exploit the system themselves.
2. The Brokerage: The vulnerability is sold to a third party, transforming the act from a potential crime into a profitable venture.
3. The Exfiltration: A separate entity handles the data theft, often leaving the vulnerability open for others to use, further distancing the original discoverer from the crime.
4. The Laundering: The final stage involves converting stolen assets or data into currency, often utilizing "work from home" mules to move funds through banking credentials before converting them back into cryptocurrency.

The Defensive Dilemma

As LLMs lower the barrier to entry for identifying vulnerabilities, the security landscape is shifting. Some observers argue that we have entered a "golden age" of net security—not because things are safer, but because the symmetry of the conflict has changed.

"Both the defense is weaker due to LLMs and attacks become stronger and cheaper. Bad combination for the rest of us."

While defenders can use AI to write patches and scan code, attackers can use the same tools to iterate on exploits at a fraction of the previous cost and time. The danger is not necessarily that AI will invent a new class of "super-exploit," but that it will democratize the ability to execute known vulnerability patterns across a wider array of targets.

Conclusion: Narrow Success vs. Broad Failure

The current state of AI in hacking can be summarized as a transition from "impossible" to "narrowly capable." LLMs are becoming genuinely proficient at bounded targets—such as smart contracts or specific API endpoints—where the rules are rigid and the scope is limited. However, they still struggle with the nuance, scale, and unpredictability of large-scale legacy systems. The real threat lies in the integration of these narrow capabilities into a broader, human-led exploit supply chain.