Trick a large language model into revealing secret passwords over multiple rounds, each with increasing safeguards.

Through prompt injection attacks, bad actors provide deceptive prompts to generative AI systems in order to manipulate their outputs and acquire sensitive data.

LLMs are trained via modeling, imitation learning, and reinforcement learning

Training large language models begins with pretraining the model to predict the next word in a sequence based on finding patterns in massive amounts of text. Patterns are then fine-tuned to model human-written dialogue and to align responses with user preferences.

Visualize how transformer architecture generates a large language model's outputs by running words and subwords assigned to specific numbers through multiple rounds of mathematical processing.

How ChatGPT learns to read

The introduction of ChatGPT last fall jumpstarted public interest in the future capabilities of artificial intelligence and what it means for society. Of many somewhat "jargony" terms thrown around, the phrase "Large Language Model" is commonly seen—but also incredibly unclear as to what it means. In the simplest sense, the programs comb through mountains of internet data in response to a question and guess at what should come next (subscription needed).

Anthropic's interpretability team discusses how AI models' thinking mimics biology

Just as humans engage in complex behaviors to survive and reproduce, teams of experts in neuroscience, virology, mathematics, and other disciplines argue that large language models develop complex mechanisms to achieve their goals. These mechanisms can be manipulated to assess how they modify outputs—much like stimulating individual neurons—to uncover how LLMs "think."

Explore rankings of large language models

The SEAL LLM Leaderboards provide benchmarks for LLMs by assessing them against agentic, frontier, and safety tasks, providing insights into each model's strengths, deficiencies, and failures.

Large language models are examples of foundation models

Such models are trained on massive amounts of data, which enable them to perform a variety of tasks, rather than being provided task-specific data to complete a narrow function. Such models come with high compute costs and potential trust issues due to the nature of their unorganized training data.

LLMs explained simply and quickly

Large language models generate responses by predicting one token at a time, based on patterns they’ve learned from massive datasets. This explainer covers their core components: data, architecture, training, and how transformer models use context to simulate conversation.

Transformer architecture can recognize and predict patterns in language

The underlying software powering text generation in AI tools associates each word or subword—called a token—with a set of values corresponding to how often it appears near other tokens. By recognizing these associations in prompts, the LLM can infer meaning.

LLMs are the backbone of AI tools that process and generate natural language

Large language models identify patterns in massive datasets of books, code, and unlabeled text from which to generate coherent responses. GPT-3 was trained on about 45 terabytes of data and uses 175 billion parameters—each parameter being a value the model adjusts as it learns.

Understanding how Claude Code and other AI coding agents function

Under human oversight, a supervising large language model interprets user tasks and delegates work to subordinate LLMs, which can generate code, fix bugs, and run tests, often most effectively for proofs of concept. Incremental backups and versioning are crucial when using such agents, which can lose details during their work as a result of compressing context history to work around memory limitations.

Watch how the Enigma machine encrypts messages and learn about its critical design flaw

Using settings selected from a series of rotors and a plugboard that let operators swap letters with any of their choosing, the machine lit up the output letter for each letter typed. While this system could generate messages requiring any of 150 quintillion potential keys, the machine's reflector never encrypted a letter as itself, providing an entry point for exploiting the system.

Generative AI tools excel at pattern recognition, not contextual accuracy

Large language models are trained on vast amounts of unstructured data, from which they develop parameters for grammar and associations between words. These connections can introduce errors due to inapplicable reasoning when used on new data in unfamiliar contexts.

Large language models struggle to play video games, despite coding them easily

As of 2025, given clear tasks, specific feedback, and test results to learn from, LLMs have been successful at coding simple video games that resemble pre-existing ones. However, given the variety of games, the lack of training datasets, and the need for spatial awareness, which LLMs find challenging, some of these models experience greater difficulty navigating and "winning" games.

Unlike the models made by large AI companies, open-weight models can have their safeguards removed or weakened, a process that has become increasingly easy and may expose users to harmful information.

Anthropic's Frontier Red Team assesses the dangers of its AI tools through 'evals'—safety tests that probe models to disclose sensitive information.

Indirect prompt injection is a major security flaw of generative AI systems

These attacks can manipulate AI behavior by hiding instructions in websites and PDFs that become part of training data. The more companies connect sites, services, and sensitive datasets to AI tools, the greater the chances of exposure to malicious code.

Claude Shannon, for whom Anthropic's AI tools are named, proved the existence of an unbreakable cipher

In the one-time pad approach, the encryption key is randomly generated and of equal length to the plaintext. This randomness ensures that every possible plaintext is equally likely—RBBHN can mean HELLO, SOLVE, CODES, and so on, using a different key—while matching the message and key lengths prevents the appearance of patterns that can be exploited.

In recent years, NIST has suggested that passwords no longer be created using a mix of symbols, numbers, and capital and lowercase letters, as was once recommended.

Neural networks are modeled after an idealization of the brain

These networks—including the backbone of ChatGPT-3, which has approximately 175 billion parameters—naturally discover the implicit rules of grammar and syntax within language by identifying patterns. Deviations in word patterns add creativity to the text generation.

How hash functions encrypt stored passwords and can be enhanced with 'salt'

Hash functions convert text into unique, random sets of characters, a process that is computationally expensive and time-consuming to undo. Because hashed outputs for many common passwords are publicly listed, some sites use a salt—additional characters appended to a password—before hashing to produce an output that is not known. When entering a password, it is salted and hashed, then checked against the database for a matching hash before permitting login.

AI models incorporate a randomness parameter when generating responses to prompts

AI models use this parameter to prevent repeated outputs by sometimes choosing less likely next words during sequential generation. However, this randomness—alongside insufficient data and training—may cause hallucinations of incorrect results.

'Context rot' explains why AI conversations degrade the longer they run

Research shows that as a large language model's context window fills up, performance declines, and models begin over-indexing on content at the start and end of a conversation while losing track of what's in the middle. The practical fix is simpler than you'd think: start a fresh chat more often.

Find out how important you are to AI tools when they generate text responses with this tool, which estimates the weights of people's names.

The top 200 most common passwords—you probably shouldn't use

Many people use the same weak passwords year after year. Check out the 2022 list of top 200 most common passwords used around the world.

Tokens—numbers assigned to words and subwords—are the basic units of data that AI tools process to generate outputs that mimic language without understanding the text itself.

Scientists increasingly can't predict what AI will do

Most AI systems are so-called "black box" models, where we only see inputs and outputs while the inner workings remain obscure. This piece, written prior to the launch of ChatGPT, cautions developers to know why and how a system produces its results, and not to sacrifice explainability for accuracy.

Model context protocol allows large language models to connect to external data

Using a unified interface, MCP simplifies how AI systems securely connect to external data, tools, and services. The protocol enables AI agents to perform tasks without ongoing human refinement or updates to manually integrate new data.

The science of cryptography, including methods to encrypt data and how to break encryption

Monoalphabetic ciphers use a set of rules to replace each letter with another consistently, making them susceptible to brute force attacks that try every kind of replacement. While more complex ciphers can transform text in less obvious ways, because certain letters are used more often than others, an analysis of letter frequency may reveal patterns that help uncover the rules that make up a cipher.

EOS and maximum length tokens tell an LLM when to stop generating a response

Tokens are words or subwords (e.g., prefixes) that a large language model processes to generate a response. During training, the model is taught that if it ever generates a special token called end-of-sequence, its response is complete.

Visualize how passkeys, which rely on asymmetric cryptography to authenticate users, work

After using biometrics or a PIN to access the private key stored on a device, it is combined with a random number generated by the site to produce a cryptographic signature. This signature is verified using the site's public key, allowing a user to log in without entering a password. This reduces the risk of fraudulent messages—phishing attacks—and data breaches.

How quantum computers can empower Shor's algorithm to break digital encryption

Secure communications largely rely on numeric keys generated by multiplying two enormous prime numbers together, and it is computationally impractical to determine these primes by trying to divide by every possible one. While Shor's algorithm can provide better guesses for this brute-force approach, it relies on a number that is difficult to determine on classical computers but much easier to calculate using the simultaneous computations enabled by quantum superposition.

Practice encrypting and decrypting text using an online library of ciphers

This tool provides users with customizable protocols, including Caesar, Affine, and Vigenère ciphers, to hide messages in transformed text and convert them back into legible formats. Mechanisms are also available to convert text into other communication languages, such as Morse code and URL encoding.

Sam Altman acknowledges the fears around creativity, copyright, and misuse of AI

In 2022, OpenAI's CEO, Sam Altman, popularized the use of large language models with the release of ChatGPT. He has emphasized the importance of involving society in shaping AI’s safety frameworks before models begin to act independently online.