- Science & Technology
- Understanding the Effects of Technology on Economics and Governance
This excerpt from the Stanford Emerging Technology Review (SETR) focuses on cryptography, one of ten key technologies studied in this educational initiative. SETR, a project of the Hoover Institution and the Stanford School of Engineering, harnesses the expertise of Stanford University’s leading science and engineering faculty. Download the full report here and subscribe here for news and updates.
The word cryptography originates from Greek words that mean “secret writing.” Once limited to simple codes and ciphers, it now relies on advanced mathematics to protect data from unauthorized access or tampering. Though largely invisible, cryptography secures many everyday interactions, from online shopping to cell phone calls.
Cryptography is essential for internet activity from messaging and banking to everyday browsing—but it cannot, on its own, guarantee the confidentiality, integrity, or availability of information. Various vulnerabilities ensure that cybersecurity will remain an ongoing challenge. These include technical vulnerabilities in the digital systems that humans operate and use; human vulnerabilities, such as the tendency to bypass security mechanisms because using them is considered inconvenient; and strong incentives for attackers.
Cryptography also is the enabling technology of blockchain, which is the enabling technology of cryptocurrencies.
Keeping it safe
Computer security traditionally focuses on safeguarding computer systems against unauthorized access and misuse. It emphasizes the core principles of confidentiality, integrity, and availability—with all three also known collectively as the CIA triad. Confidentiality refers to the privacy of data (i.e., preventing unauthorized disclosure). Integrity refers to preserving data (i.e., guarding against unauthorized alterations). Availability refers to data and resources being accessible to authorized users, especially during critical times.
Historically, computer security focused on protecting individual machines from actions perpetrated by malicious actors, whether individuals or states. Over time, the focus has expanded—first to securing the infrastructure of increasingly networked systems and now to addressing vulnerabilities in machine learning (ML) models. Cryptography is one of many tools whose use can enhance computer security. However, the protections afforded by even perfect cryptography can often be circumvented by taking advantage of vulnerabilities in the computer systems on which that cryptography is implemented. For example, if an intercepted encrypted message is too hard to decrypt, the attacker’s focus will most likely be on exploiting vulnerabilities in computer security to obtain the message before the sender encrypts it or after the receiver decrypts it.
Thus, cryptography and security are inseparable, but using cryptography is by no means a guarantee of security.
As ML systems move into high-stakes settings—including autonomous vehicles, financial platforms, and health care diagnostics—their security under adversarial conditions is becoming a critical concern. In an ML system, small, malicious changes to inputs can cause large, unexpected model failures. These brittle responses undermine trust in a system’s ability to operate safely in environments where reliability is paramount.
This fragility stems from a core asymmetry. While ML performs well on the inputs that most average users would give it, it often fails on inputs that are crafted by deliberately malicious adversaries. In other words: ML systems are great for random data, but they often perform poorly when confronted with deliberately crafted adversarial data.
As AI agents—autonomous AI soft-ware programs that have access to important data or controls—become increasingly popular, there is a growing risk that these could be “tricked” by malicious content on the internet.
No current approach offers broad protection across all tasks, data types, or adversarial techniques. The field remains in an arms race: new attacks emerge rapidly, while robust, scalable defenses continue to lag behind. In this landscape, any claim to deploy ML to solve a problem should prompt an immediate question: what have you done about adversarial inputs and attacks?
The quantum future
When realized, quantum computing will pose a significant threat to today’s public-key algorithms. Experts disagree on how long it will take to build quantum computers that are capable of this, but under the May 2022 National Security Memorandum 10, Promoting US Leadership in Quantum Computing While Mitigating Risks to Vulnerable Cryptographic Systems, the US government has initiated the transition to quantum-resistant public-key algorithms. Many experts in the field expect quantum-resistant algorithms will be widely available by the time quantum computing comes online.
At the intersection of quantum computing and cryptography are two important issues: (1) that support for the transition to a quantum-resistant encryption environment should continue with urgency and focus, and (2) that messages protected by pre-quantum cryptography will be vulnerable in a post-quantum world.
If those messages have been saved by adversaries (which is likely in the case of parties like Russia), those bad actors will be able to read a host of old messages. Containing secrets from the past, they may reveal embarrassments and dangers with potentially detrimental policy implications.
