Artificial intelligence (AI) forces us to confront this question. It is a technology that in theory amplifies the power of its users: A manager, marketer, political campaigner, or opinionated internet user can utter a single instruction, and see their message—whatever it is—instantly written, personalized, and propagated via email, text, social, or other channels to thousands of people within their organization, or millions around the world. It also allows us to individualize solicitations for political donations, elaborate a grievance into a well-articulated policy position, or tailor a persuasive argument to an identity group, or even a single person...

Abstract: We present evidence that adversarial poetry functions as a universal single-turn jailbreak technique for Large Language Models (LLMs). Across 25 frontier proprietary and open-weight models, curated poetic prompts yielded high attack-success rates (ASR), with some providers exceeding 90%. Mapping prompts to MLCommons and EU CoP risk taxonomies shows that poetic attacks transfer across CBRN, manipulation, cyber-offence, and loss-of-control domains. Converting 1,200 ML-Commons harmful prompts into verse via a standardized meta-prompt produced ASRs up to 18 times higher than their prose baselines. Outputs are evaluated using an ensemble of 3 open-weight LLM judges, whose binary safety assessments were validated on a stratified human-labeled subset. Poetic framing achieved an average jailbreak success rate of 62% for hand-crafted poems and approximately 43% for meta-prompt conversions (compared to non-poetic baselines), substantially outperforming non-poetic baselines and revealing a systematic vulnerability across model families and safety training approaches. These findings demonstrate that stylistic variation alone can circumvent contemporary safety mechanisms, suggesting fundamental limitations in current alignment methods and evaluation protocols...

We have just published the book Rewiring Democracy: How AI will Transform Politics, Government, and Citizenship. In it, we take a clear-eyed view of how AI is undermining confidence in our information ecosystem, how the use of biased AI can harm constituents of democracies and how elected officials with authoritarian tendencies can use it to consolidate power. But we also give positive examples of how AI is transforming democratic governance and politics for the better...

In 2008, social media was just emerging into the mainstream. Facebook reached 100 million users that summer. And a singular candidate was integrating social media into his political campaign: Barack Obama. His campaign’s use of social media was so bracingly innovative, so impactful, that it was viewed by journalist David Talbot and others as the strategy that enabled the first term Senator to win the White House...

For example, in a medical context, if an AI was demonstrably better at reading a test result and diagnosing cancer than a human, you would take the AI in a second. You want the more accurate tool. But justice is harder because justice is inherently a human quality in a way that “Is this tumor cancerous?” is not. That’s a fact-based question. “What’s the right thing to do here?” is a human-based question...

A new attack called ‘CometJacking’ exploits URL parameters to pass to Perplexity’s Comet AI browser hidden instructions that allow access to sensitive data from connected services, like email and calendar.

In a realistic scenario, no credentials or user interaction are required and a threat actor can leverage the attack by simply exposing a maliciously crafted URL to targeted users.

[…]

CometJacking is a prompt-injection attack where the query string processed by the Comet AI browser contains malicious instructions added using the ‘collection’ parameter of the URL...

As authoritarianism is rising around the world, AI-generated “slop” is overwhelming legitimate media, while AI-generated deepfakes are spreading misinformation and parroting extremist messages. AI is making warfare more precise and deadly amidst intransigent conflicts. AI companies are exploiting people in the global South who work as data labelers, and profiting from content creators worldwide by using their work without license or compensation. The industry is also affecting an already-roiling climate with its ...

This system assigns action items and determines the importance of what is said. If it becomes necessary to revisit the facts of the meeting, its summary is treated as impartial evidence.

But clever meeting attendees can manipulate this system’s record by speaking more to what the underlying AI weights for summarization and importance than to their colleagues. As a result, you can expect some meeting attendees to use language more likely to be captured in summaries, timing their interventions strategically, repeating key points, and employing formulaic phrasing that AI models are more likely to pick up on. Welcome to the world of AI summarization optimization (AISO)...

Below, co-authors Bruce Schneier and Nathan E. Sanders share five key insights from their new book, Rewiring Democracy: How AI Will Transform Our Politics, Government, and Citizenship.

What’s the big idea?

AI can be used both for and against the public interest within democracies. It is already being used in the governing of nations around the world, and there is no escaping its continued use in the future by leaders, policy makers, and legal enforcers. How we wire AI into democracy today will determine if it becomes a tool of oppression or empowerment...

Many decades ago, U.S. Air Force Colonel John Boyd introduced the concept of the “OODA loop,” for Observe, Orient, Decide, and Act. These are the four steps of real-time continuous decision-making. Boyd developed it for fighter pilots, but it’s long been applied in artificial intelligence (AI) and robotics. An AI agent, like a pilot, executes the loop over and over, accomplishing its goals iteratively within an ever-changing environment. This is Anthropic’s definition: “Agents are models using tools in a loop.”...

Over the summer, hackers proved the concept, industry institutionalized it, and criminals operationalized it. In June, AI company XBOW took the top spot on HackerOne’s US leaderboard after submitting over 1,000 new vulnerabilities in just a few months. In August, the seven teams competing in DARPA’s AI Cyber Challenge ...

The widespread fear that AI would be used to manipulate the 2024 US election seems rather quaint in a year where the president posts AI-generated images of himself as the pope on official White House accounts. But AI is a lot more than an information manipulator. It’s also emerging as a politicized issue. Political first-movers are adopting the technology, and that’s opening a ...

Abstract: Large Language Model (LLM)-enabled agents are rapidly emerging across a wide range of applications, but their deployment introduces vulnerabilities with security implications. While prior work has examined prompt-based attacks (e.g., prompt injection) and data-oriented threats (e.g., data exfiltration), time-of-check to time-of-use (TOCTOU) remain largely unexplored in this context. TOCTOU arises when an agent validates external state (e.g., a file or API response) that is later modified before use, enabling practical attacks such as malicious configuration swaps or payload injection. In this work, we present the first study of TOCTOU vulnerabilities in LLM-enabled agents. We introduce TOCTOU-Bench, a benchmark with 66 realistic user tasks designed to evaluate this class of vulnerabilities. As countermeasures, we adapt detection and mitigation techniques from systems security to this setting and propose prompt rewriting, state integrity monitoring, and tool-fusing. Our study highlights challenges unique to agentic workflows, where we achieve up to 25% detection accuracy using automated detection methods, a 3% decrease in vulnerable plan generation, and a 95% reduction in the attack window. When combining all three approaches, we reduce the TOCTOU vulnerabilities from an executed trajectory from 12% to 8%. Our findings open a new research direction at the intersection of AI safety and systems security...

To most on the American left, the DOGE end game is a dystopic vision of a government run by machines that benefits an elite few at the expense of the people. It includes AI ...

“Invitation Is All You Need! Promptware Attacks Against LLM-Powered Assistants in Production Are Practical and Dangerous”

Abstract: The growing integration of LLMs into applications has introduced new security risks, notably known as Promptware­—maliciously engineered prompts designed to manipulate LLMs to compromise the CIA triad of these applications. While prior research warned about a potential shift in the threat landscape for LLM-powered applications, the risk posed by Promptware is frequently perceived as low. In this paper, we investigate the risk Promptware poses to users of Gemini-powered assistants (web application, mobile application, and Google Assistant). We propose a novel Threat Analysis and Risk Assessment (TARA) framework to assess Promptware risks for end users. Our analysis focuses on a new variant of Promptware called Targeted Promptware Attacks, which leverage indirect prompt injection via common user interactions such as emails, calendar invitations, and shared documents. We demonstrate 14 attack scenarios applied against Gemini-powered assistants across five identified threat classes: Short-term Context Poisoning, Permanent Memory Poisoning, Tool Misuse, Automatic Agent Invocation, and Automatic App Invocation. These attacks highlight both digital and physical consequences, including spamming, phishing, disinformation campaigns, data exfiltration, unapproved user video streaming, and control of home automation devices. We reveal Promptware’s potential for on-device lateral movement, escaping the boundaries of the LLM-powered application, to trigger malicious actions using a device’s applications. Our TARA reveals that 73% of the analyzed threats pose High-Critical risk to end users. We discuss mitigations and reassess the risk (in response to deployed mitigations) and show that the risk could be reduced significantly to Very Low-Medium. We disclosed our findings to Google, which deployed dedicated mitigations...

Bargury’s attack starts with a poisoned document, which is shared to a potential victim’s Google Drive. (Bargury says a victim could have also uploaded a compromised file to their own account.) It looks like an official document on company meeting policies. But inside the document, Bargury hid a 300-word malicious prompt that contains instructions for ChatGPT. The prompt is written in white text in a size-one font, something that a human is unlikely to see but a machine will still read.

In a proof of concept video of the attack...