OpenAI recently penned an open letter calling for a coordinated worldwide effort against AI-enabled cyberattacks. The letter featured more than 100 day-one signatories hailing from across the global technology landscape, illustrating the scope of the threat: Generative AI is compressing the time between when a security flaw goes live and when a bad actor can exploit it. A growing body of research reveals a lack of enterprise awareness about AI security protocols. One broker reports that, across recent conversations with customers and partners, none knew how their firm secures its AI agents, despite all of them having agents in action.
Yet AI also carries cybersecurity benefits, giving firms new tools to automate threat detection and response. Against this backdrop, the market is moving away from tools and platforms that wait for a human to approve each step in the workflow, and toward context-aware ones that act autonomously.
Discover the security risks and opportunities that genAI creates for firms in this article, powered by insights from AlphaSense’s SuperAnalyst.
AI-Enabled Cyberattacks Accelerate, At a Cost
AI has lowered the technical skill needed to build and carry out an effective cyberattack, while also reducing the cost and time needed to prepare an attack. AI-driven attacks are on the rise: Roughly three-quarters of detected malware now shows signs of being able to evolve dynamically on its own due to AI, according to broker research in AlphaSense. AI-assisted cyberattacks are especially costly, averaging $5.7 million per attack in 2025.
Meanwhile, frontier AI models are proving difficult to contain, even in controlled environments. In July 2026, OpenAI disclosed that two of its models, tested with reduced cyber refusals, escaped a sandboxed evaluation, exploited a previously unknown vulnerability, and breached Hugging Face's production infrastructure — reportedly to retrieve the answer key for the benchmark they were being graded on. Hugging Face detected and contained the intrusion on July 16, five days before OpenAI publicly attributed it to its own testing. Days later, Anthropic disclosed that a misconfiguration in a third-party cybersecurity evaluation left several Claude models connected to the open internet. While pursuing a simulated capture-the-flag task, the models breached the real systems of three organizations. Anthropic said the models had been told they had no internet access, treated the real systems as part of the exercise, and did not deliberately try to escape. Both cases show that today’s most advanced models can find and follow an attack path on their own.
OpenAI responded by reallocating roughly a quarter of its production engineers to security functions. Co-founder Greg Brockman said that AI makes it easier to chain small flaws into a real attack, so flaws teams once tolerated now need fixing. Anthropic CEO Dario Amodei shares that viewpoint, saying the industry has to “slow the pace at which we improve the capabilities of AI models” so the safety work can keep up with what the newest models are capable of do. Sam Altman gave a similar warning at the UN Security Council: The moment “calls for extreme care,” and the industry “must not accept too much technological risk just because the benefits are too great.” As the models gain autonomy, keeping human control matters more, and letting them improve themselves without safeguards introduces unacceptable risk, Altman added.
Security concerns have the potential to hinder AI adoption. Findings from our 2026 State of AI for Business and Finance report reveal that security is the top organizational barrier to further genAI investment. This finding holds true across regions but is especially pronounced in Asia-Pacific, where more than half of respondents agree security is preventing broader adoption. Highly regulated end markets such as banking and financial services, healthcare, and government are particularly conservative on deployment and continue to face adoption hurdles.
AI Boosts Cyber Defense Capabilities and ROI
On the opposite end of the spectrum, AI is providing the same opportunities for cyber defense firms. For example, security teams can now use AI to understand network traffic and user behavior and flag anomalies in real time. Because these systems understand behavioral intent, they can detect threats faster than relying on traditional signals. Teams also have access to “adaptive AI” that understands a company’s internal systems. Hugging Face’s response to the OpenAI incident shows the point in practice: Its security team used its own AI systems, including open-source models, to contain the intrusion and reconstruct the attack path. It was initially blocked when commercial frontier models refused to analyze real exploit payloads.
These boosted capabilities are already translating into cost savings for firms: Broker research shows that companies using AI and automation extensively across their security work contain breaches 65 days sooner and spend nearly $2 million less per breach, compared with peers. Yet the same research reveals that only 36% of firms use these tools across the full lifecycle.
The composition of defense budgets is also shifting, with a growing share going specifically toward shoring up internal AI. One sell-side survey revealed that 7.6% of cyber budgets are now allocated to securing AI, up a percentage point from last year. Addressing this issue is a point of emphasis among business leaders: 40% of surveyed CIOs called security the most overlooked line in their firm’s AI budget.
The Vendor Response: Bundled Offerings
As AI agents become widespread across enterprise applications, the scope of a potential attack is widening, creating a new class of risks to monitor. Rather than using a separate tool to address each individual risk, enterprise buyers increasingly are looking for a single vendor that secures AI workloads, machine identities, and autonomous agents at once.
Palo Alto Networks is considered a major beneficiary of this shift: The platform combines next-gen firewalls, Prisma Cloud, and Cortex XSIAM into a single bundled offering. Meanwhile, Zscaler management recently said 70% of its security for AI deals last quarter also included its data-security solution, indicating increased appetite for bundled solutions. Crowdstrike, which calls its Falcon platform “cybersecurity's infrastructure layer for AI adoption,” offers agentic solutions spanning endpoint, next-gen SIEM, and identity.
At the same time, multi-vendor partnerships remain key. Okta recently formed the Blueprint Alliance, a multi-vendor open architecture for discovering agents, controlling permissions, and shutting them down when they misbehave. Okta’s key partners in this effort are AWS, CrowdStrike, Databricks, Google Cloud, Salesforce, ServiceNow, and Zscaler.
Identity Is the Biggest Near-Term Risk, and Opportunity
AI agents move quickly and often run with broad standing permissions by default that experts say can be “pretty crazy” in scope, meaning that effective security depends on tight data controls. Yet today, CISOs note that “no one really has the right infrastructure” in place to ensure that an agent is authorized for what it is about to do, which concentrates risk at the identity layer where permissions are determined.
No complete solution to the identity issue exists today, and experts expect the race to shape the next wave of investment. Key variables to watch near term include whether identity-first governance can keep pace with autonomous agents, which vendors consolidate the control layer, and how quickly regulators move on new AI security standards. But for now, the technology is considered to be advancing faster than firms can keep up.
“[T]rue autonomous execution is still very, very early. Identity for agents and authorization and governance and observability, data access, determining how much authority you give an agent, all that stuff is still being worked out. The interesting thing is that the models themselves are probably ahead of enterprise adoption. The technology is moving faster than companies can really design their processes and controls around it.”
– Chief Information Security Officer at Mastech Digital, September 2026 Call
Eliminate Blind Spots With Always-On Intelligence
AI has reduced the technical skill needed to conduct a cyberattack. Staying ahead of these threats means seeing how they’re being discussed in real time, across earnings calls, broker research, and expert commentary.
Maintaining that kind of coverage manually requires more time and manpower than any team can afford. Enter AlphaSense's SuperAnalyst, an always-on AI agent that plans, monitors, executes, and delivers entire workflows on your behalf. SuperAnalyst is grounded in AlphaSense’s industry-leading library of premium, proprietary content, meaning every finding is grounded in a verifiable source.
Find out what AlphaSense can do for your team. Start your free 2-week trial today.
