We are undoubtedly in the middle of an era defined by artificial intelligence (AI), and companies across multiple industries are investing heavily in their AI capabilities. According to these firms, one of the most promising avenues in the AI industry is AI agents, but so far, they have fallen far short of expectations. Ironically, it seems like their most effective use is finding and exploiting vulnerabilities in various crypto projects.

Below, we’ll discuss the current state of the AI agent market and how they haven’t quite lived up to the hype yet.

Companies Struggle to Put AI Agents to Work

AI agents are programs powered by large language models (LLMs), designed to plan and execute tasks autonomously. While many saw them as a big game changer, they haven’t lived up to the promise, in part because of the high and misaligned expectations, as the current AI agents are not fully prepared to handle complex tasks.

Moreover, integrating AI agents with existing systems remains a challenge, as is interaction between different AI agents. Many companies have also been struggling to figure out how AI agents fit into their business goals. The issues of data safety and privacy have also been major concerns among these firms as they deal with a lot of sensitive user data.

Finally, like every emerging technology, many companies are in a wait and watch mode and studying how these agents perform. This includes the cost-benefit analysis, as for some companies that have access to cheap labor, manual agents might still be a better bet financially.

Existing AI Agents Don’t Meet the Hype

It seems like staying on the sidelines for now has been a good idea as researchers at Carnegie Mellon University released a paper in May that shows that Google’s Gemini 2.5 Pro, which was otherwise the best-performing AI agent, couldn’t fully complete real-world office tasks a whopping 70% of the time. If that sounds too high, the research showed that OpenAI’s GPT-4o had a failure rate of 91.4% while it stood at 92.6% for Meta’s Platforms’ Llama-3.1-405b. In short, they just don’t work very well.

While markets are still quite bullish on AI, Gartner predicts that 40% of the current AI agent contracts will get cancelled by 2027. “Agent washing,” or the process of labeling existing systems with little to no improvement as AI agents, is not helping matters either. Gartner estimates that of the thousands of AI agents deployed by companies, just about 130 are real.

“Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied,” said Anushree Verma, senior director analyst at Gartner.

She added, “Most agentic AI propositions lack significant value or return on investment (ROI), as current models don’t have the maturity and agency to autonomously achieve complex business goals or follow nuanced instructions over time.”

AI Agent Can Help Detect Crypto Vulnerability

While AI companies are struggling to make AI agents into a functional, profitable product, researchers at the University of Sydney (USYD) in Australia and University College London (UCL) have devised an AI agent named A1 that can discover and exploit vulnerabilities in blockchain smart contracts.

These contracts can have bugs in their code which can be exploited to steal money. The crypto industry is infamous for such hacks and last year lost $1.5 billion to hacking activity bringing the cumulative total to a mammoth $11.5 billion since 2017, according to data from Web3 security platform vendor Immunefi.

The A1 agent system was developed by Liyi Zhou, a lecturer in computer science at USYD, and Arthur Gervais, a professor in information security at UCL. It uses AI models from US-based OpenAI and Google, as well as models from DeepSeek (the Chinese AI startup that created waves with its low-cost model) and Alibaba.

While AI agents are notorious for phantom flaws, A1 seems to be a lot more reliable. It demonstrated a success rate of nearly 63% on the Verite benchmark when tested on 36 real-world vulnerable contracts on Binance Smart Chain blockchains and Ethereum.

In an email to The Register, Zhou said, “A1 performs full exploit generation.” He added, “This is important. This is unlike other LLM security tools. The output is not just a report, but actual executable code. A1 is really close to a human hacker.”

Could the A1 Agent Be Profitable?

The A1 agent could theoretically make a profit as it can earn more from spotting the exploits in smart contracts than it takes to operate it. Naturally, it would be unethical and illegal to perform such exploits, but it could likely turn a profit as a white-hat hacker.

“To give a concrete example [from the paper], Figure 5 shows that o3-pro remains profitable even if only 1 out of every 1000 scans leads to a real vulnerability – as long as that vulnerability was introduced in the last 30 days,” explains Zhou in the paper.

While the draft paper said that A1 will be released as open source, Zhou ruled out that possibility as the AI agent could be misused by criminals.

“We’ve removed the mention of open source (arXiv will show tomorrow) as we’re not yet sure whether it’s the right move, given how powerful A1 is,” said Zhou.

The Artificial Intelligence Rally Is Pretty Much Back Even as Companies Scramble for Profitability

Despite the AI agent struggle, the AI market is still booming. Recently, Nvidia – whose GPUs power the AI ambitions of other companies – became the first company ever to reach a market cap of $4 trillion.

Other AI plays in the Big Tech space have shown mixed results, though. For instance, Facebook parent Meta Platforms cited higher AI investments and raised its 2025 capex budget to between $64 billion and $72 billion as compared to the previous guidance of $60 billion to $65 billion.

In its Q1 2025 earnings call, Meta said that its AI assistant Meta AI has surpassed 1 billion active users. The company has set up its Superintelligence Labs to lead its AI efforts, and the segment will be co-led by Alexandr Wang, whom Meta hired as part of its 49% acquisition of Scale AI that he founded.

In its March quarter earnings call, Microsoft said that GitHub Copilot assistant user count has swelled over fourfold over the last year to 15 million. Tesla, which has positioned itself as an AI company amid falling vehicle sales, has also rolled out its robotaxi service in Austin.

Nvidia’s Profits Have Swelled Amid High Demand for GPUs

Meanwhile, while AI companies have made incremental progress, not many, barring the notable exception of Nvidia and other chipmakers, have much concrete to show in terms of profitability. As a McKinsey report aptly puts it, “Gen AI is everywhere—except in company P&L.”

That said, AI is indeed seeing at least moderate growth in company toplines. OpenAI, for instance, has hit an annualized revenue run rate of $10 billion, as compared to $5.5 billion in 2025.

Overall, these are still early days for AI – and by extension AI agents – and one reason the technology’s progress has disappointed some is because of the sky-high expectations. However, as the technology gets better with time, we will almost certainly see AI agents perform much better than they are currently.