100+ Tech Firms Warn of an AI Cyberattack Surge

What happened?

OpenAI published an open letter on August 27 calling for collective action on cyber defense. More than 100 organizations signed it, including major AI companies, cloud and software providers, cybersecurity firms, financial organizations and internet infrastructure companies. OpenAI's published signatory list includes Anthropic, Google, Microsoft, AWS, Cisco, Cloudflare, CrowdStrike, IBM, Oracle, Palo Alto Networks and many others.

The timing matters. The letter is not saying that every AI system is currently carrying out attacks against critical infrastructure. It is a warning about the direction of travel: more capable models can give attackers more ability to automate parts of reconnaissance, vulnerability research, coding and other steps in an intrusion.

At the same time, the signatories argue that AI can improve defense. That makes the letter less about banning the technology and more about closing the gap between offensive capability and defensive readiness.

What does the letter ask for?

The letter is aimed at four groups: every organization, cybersecurity companies and technology partners, governments, and frontier AI companies.

One of the more important ideas in the letter is that security teams should use AI too. The signatories argue that capable models can help defenders find weaknesses, improve coverage and respond faster, especially when security teams are already short on staff and resources.

Is the threat already real?

There is independent evidence that AI is being used in cyber operations, although that does not mean the specific scenarios in the letter have already happened at large scale.

On August 19, the U.S. National Security Agency and partner agencies said threat actors were conducting reconnaissance and capability development against U.S.-based Siemens programmable logic controllers using AI-generated exploitation scripts disguised as legitimate monitoring tools. The advisory covered sectors such as manufacturing, energy, water and wastewater, chemicals, agriculture and other critical facilities.

That distinction is important. The official advisory supports the claim that AI-generated exploit code is already part of observed threat activity. It does not prove that an AI agent has independently taken over a power grid or water system. Good security reporting should keep those claims separate.

Why AI changes cyber risk

AI can compress the time needed for parts of an attack chain. A capable model can assist with code generation, log review, documentation, language translation, data analysis and repetitive task execution. The value to an attacker is not always one “super-hacker” model. It can be the ability to scale many small tasks.

That creates a practical problem for defenders: security programs are often built around human review, limited staffing and patch cycles that take time. Attackers may use automation to test more ideas, generate more variants and adapt faster.

There is also a second risk. AI-generated code can enter normal software development. The open letter therefore calls attention to the security of AI-generated code rather than treating the model itself as the only risk surface.

Agentic systems add another layer. When a model can call tools, access services or take actions, organizations need strong identity, permissions, logging and monitoring around the agent as well as the underlying model.

Who is affected?

Critical infrastructure is a clear concern because a compromise can affect physical operations, not just data. The NSA's recent PLC advisory shows why industrial environments deserve attention.

But the issue is much wider. Software companies, banks, cloud providers, hospitals, online services and government agencies all depend on systems that can be probed, misconfigured or exploited. Smaller organizations may face an even harder problem if they do not have large security teams.

For technology leaders, the main question is becoming less “Are we using AI?” and more “Where does AI have access to systems, data, credentials and code, and what controls sit around those access paths?”

What organizations should do now

1. Fix the highest-risk exposure first

Do not start with a generic AI policy document. Start with internet-facing systems, known high-risk vulnerabilities, privileged accounts and critical services. The open letter explicitly puts urgent remediation ahead of routine activity.

2. Review AI-generated code before it reaches production

AI can make developers faster, but generated code still needs normal security review, testing and dependency checks. Treat model output as untrusted until it passes the same engineering controls used for human-written code.

3. Strengthen identity and least privilege

When an AI agent can call APIs or tools, give it only the permissions needed for the task. Use separate identities, short-lived credentials where practical, and logs that show which agent or service made each action.

4. Test defenses against AI-assisted attacks

Run realistic security testing rather than assuming existing controls are enough. The collective letter specifically calls for continuous testing against advanced cyber capabilities.

5. Protect industrial and critical systems

For operational technology, keep control systems away from direct internet exposure when possible, use segmentation and strong access controls, and follow the mitigations in current government advisories. The FBI and EPA have also warned about internet-facing PLC attacks in the water and wastewater sector.

6. Improve monitoring around AI systems

Track model use, tool calls, unusual access patterns and changes made by automated systems. The more power an AI system has, the more important its audit trail becomes.

What this means for AI and technology teams

This is not only a cybersecurity story. It is also a sign that AI infrastructure, model capability and security architecture are becoming tightly linked.

The recent GLM-5.3-Flash analysis on ToolboxKart looked at how model architecture and inference efficiency affect real workloads. The recent OpenAI Jalapeño chip analysis looked at the hardware side. This cyber-defense warning adds another layer: the systems around capable models may become just as important as the models themselves.

For engineering and platform teams, security should be part of AI deployment design from the start. That includes permissions, network controls, secrets handling, logging, model gateways and review of generated code.

For teams that manage websites and digital services, the same principle applies at a simpler level: keep public systems patched, limit unnecessary exposure, review HTTP security controls and make sure automated tools cannot gain permissions they do not need. ToolboxKart's HTTP Header Checker can help review the response headers exposed by a public web service.

What is still unclear?

The biggest uncertainty is scale. The letter predicts that AI-enabled attacks will become more widespread and sophisticated, but it does not provide a single forecast for how many attacks will occur, how quickly costs will fall for attackers, or which sectors will be hit hardest.

It is also important not to turn a public warning into proof that every AI-enabled threat is autonomous. Many real attacks can remain partly human-driven while using AI for selected tasks.

Another open question is how quickly defensive tools improve. The signatories believe defenders have an opportunity to use AI to strengthen security now. Whether that defensive advantage is achieved will depend on deployment quality, staffing, access to data, testing and basic security hygiene.

Why this matters

The strongest part of this story is the unusual level of agreement. Direct competitors in AI, cloud, cybersecurity and finance are all making the same broad argument: faster AI capability needs a faster defensive response.

That does not make every prediction in the letter certain. It does, however, make the issue hard to dismiss as a single company's product message.

The practical takeaway is simple. Organizations do not need to wait for a dramatic AI attack to improve security. The useful work is already known: reduce exposed systems, fix serious weaknesses, limit permissions, test defenses, protect critical infrastructure and monitor AI-driven actions.