0:08
AI Models Overthink Problems—and It’s a Security Risk
Large language models (LLMs) that can think through problems step-by-step have significantly increased the scope of tasks that AI can tackle. But new research suggests these reasoning capabilities also introduce a critical vulnerability that could allow attackers to slow these systems to a crawl.
While earlier generations of LLMs would immediately produce a response to a user’s request, today’s most advanced models generate an internal monologue where they break down the problem into steps and reason about the best way to tackle it before providing an answer. This has allowed AI to tackle increasingly complex problems, particularly in areas like coding and math.
However, previous research has shown that these models are susceptible to sometimes producing excessively long streams of reasoning that do little to boost performance, a phenomenon known as “overthinking.” In research presented this week at the International Conference on Machine Learning 2026 in Seoul, researchers from Zhejiang University and e-commerce giant Alibaba in China demonstrate that they can deliberately induce overthinking by subjecting models to logically inconsistent prompts. The result is a form of denial-of-service attack on commercial AI models.
The team has developed an evolutionary algorithm that corrupts the logical structure of prompts, causing models to spiral into overthinking as they attempt to reason through fundamentally unsolvable problems. Generating longer responses costs more and increases the load on a model provider’s servers, so if done at scale, the researchers say, this could significantly degrade the experience of legitimate users. The attack was effective against reasoning models from leading AI companies including DeepSeek-R1, Alibaba’s Qwen3-Thinking, OpenAI’s GPT-o3, and Google’s Gemini 2.5 Flash and resulted in outputs up to 26 times as long as standard responses on a standard math benchmark.
“Across multiple datasets and reasoning models, our method substantially amplifies the output length,” Wei Cao, a masters student at Zhejiang University, wrote in an email to IEEE Spectrum. “Our results suggest that overthinking is not an isolated phenomenon specific to individual models, but rather a shared vulnerability among modern reasoning models.”
The team’s approach builds on previous research from another group of researchers that showed reasoning models tend to overthink when faced with a question in which a key premise has been removed—such as asking how far someone who walks ten miles a day covers in total without specifying how many days they walked for. Rather than identifying that the problem is unsolvable, models often engage in extended but ultimately fruitless reasoning loops in an attempt to answer the question.
Taking the idea a step further, the authors took 940 problems from three math benchmark datasets and used an LLM to break down their logical structure into a set of premises and a final question. The genetic algorithm then jumbled these up using a variety of “mutations,” including swapping premises between problems, adding extra premises to problems, deleting existing premises from problems, and swapping the final questions between two sets of premises.
After each round of mutations, the problems are scored on how many words they cause a target model to output and also whether they increase the frequency of specific linguistic markers of overthinking—words like “but,” “wait,” “maybe,” or “alternatively.” The problems that scored highest on both measures are retained and the remaining ones are jumbled up again, and this process is repeated for five generations. Crucially, the approach doesn’t require access to the internals of a model and can generate malicious prompts by simply querying the target, which makes it possible to attack closed-source commercial services, says Cao.
The researchers found that the approach consistently led to outputs several times longer than those generated by the unmodified questions for the reasoning models they tested it on. The biggest jump came from DeepSeek-R1 on the MATH dataset, which is made up of problems from high school math competitions, where the maximum output was 26.1 times as long as the longest response the model provided to unaltered questions. While the main thrust of the research was focused on math problems, the authors also tested it on coding, scientific reasoning, and dialogue challenges, and observed significant jumps in output length in all three.
One challenge for the approach is that developing the malicious prompts requires repeated queries to expensive reasoning models, which Cao admitted could limit its cost-effectiveness. However, the researchers also demonstrated that when they used a smaller, cheaper model to generate the malicious prompts they were still able to induce the target models to produce outputs several times longer than normal.
4:02
OpenAI's GPT-5.6 launches Thursday after a delay forced by the U.S. government
OpenAI slipped the new GPT‑5.6 out Thursday because the U.S. regulators finally gave the green light after a round of extra safety checks. The hold‑up wasn’t about the model’s size or a new architecture so much as the extra testing pipeline the agency demanded—something that forced OpenAI to tweak its internal approval flow before they could ship.
What’s interesting is how the model handles code now. In head‑to‑head runs against Anthropic’s Claude Mythos 5, GPT‑5.6 not only outperformed it on the standard coding benchmarks, but it did so while consuming roughly half the compute budget. That translates to a noticeable cost dip for developers who run the model at scale.
So, after the bureaucratic pause, the rollout lands with a tighter, cheaper coding edge, and the community finally gets to see what those extra tests actually nudged under the hood.
4:48
Muse Image is technically impressive, but Meta's use of Instagram photos raises questions
Meta’s new Muse Image is an agent‑style generator, so it doesn’t just spit out a picture—it can run code, pull data from the web, and iterate on its own output. That loop of searching, tweaking and re‑rendering feels a lot like what OpenAI did with GPT‑4’s image side, but Meta’s version leans heavily on its own internal tools, making the whole process surprisingly self‑contained.
What’s catching people’s eyes is the @‑mention trick: you can tag a name and Muse will scrape publicly available Instagram posts to stitch together a portrait of that person, even if they never gave permission. Meta says you can opt out, but the default is “use it if it’s public,” which feels like a legal gray area.
Because the model pulls from real‑world feeds, regulators are already flagging it. In Europe the GDPR and the upcoming AI Act both have clauses that could deem this kind of implicit data harvesting non‑compliant, especially if users can’t easily withdraw consent.
So, technically it’s a neat step forward, but the way it leans on Instagram content is sparking a bigger conversation about privacy, consent, and where the line should be drawn for AI‑generated media.
5:49
Meta tests always-on AI glasses that capture your entire day
Meta is testing a prototype of AI-powered headsets with "Super Sensing" that use a camera and microphone to record every moment of the wearer's life.
The article Meta tests always-on AI glasses that capture your entire day appeared first on The Decoder.
6:06
The Threshold Is a Price, Not a Percentage
How to decide when an AI agent should act on its own by using cost asymmetry instead of a fixed confidence cutoff
The post The Threshold Is a Price, Not a Percentage appeared first on Towards Data Science.
6:20
Intelligence is Free, Now What? Data Systems for, of, and by Agents
The thing that’s been shaking up everything isn’t a new model architecture, it’s the price tag. A GPT‑4‑scale run that cost about thirty bucks per million tokens a year ago now slides under a dollar, and some services are already flirting with ten cents. Across the board inference prices are shrinking anywhere from ten‑fold to nine‑hundred‑fold each year, so the economics of “just‑ask‑the‑model” are finally becoming trivial.
Because the compute bill is disappearing, the bottleneck moves to the way we feed those agents data. Imagine a swarm of agents spawned for a single user query, each probing a different slice of a database. In practice 80‑90 % of the sub‑queries end up being duplicates, so the system ends up doing a lot of wasted work. The trick is to treat the database as an active partner: cache overlapping results, share scans, and even serve approximate answers fast enough for the agents to decide whether to dig deeper.
That shift also means we need a new substrate for the agents themselves. Storing everything in markdown files and grepping through them works when you have a handful of bots, but it collapses once you have thousands juggling long‑running tasks. What’s emerging is a structured memory layer—think a lightweight, queryable store that can surface just the bits an agent needs without flooding its context window.
Finally, the most interesting part is that agents are now capable of building their own data pipelines on the fly. They can spin up custom views, write macros, or even generate whole ETL scripts, but we still have to figure out how to verify those self‑assembled systems. The challenge is to give the agents enough freedom to compose the right plumbing while keeping the overall behavior trustworthy. All of this is happening because the cost of raw intelligence has slipped to almost zero, and the real work now lives in how we orchestrate and remember it.
7:55
Copilot goes cheap as Microsoft phases out OpenAI and Anthropic models to cut costs
So Microsoft is swapping out the AI models from OpenAI and Anthropic in some of their products, like Excel and Outlook, for their own models, called MAI. This change is already happening, with a huge number of queries going through these new models every week. The goal here is to cut costs, and the head of AI at Microsoft wants to get to a point where they don't have to pay for external models at all. For people using Copilot, this might mean the performance isn't as strong as it used to be, but the price will stay the same.
It's interesting because it shows how Microsoft is trying to balance the cost of using these external models with the benefits they provide. By developing their own models, they can save money in the long run, but it's not clear if these new models will be as effective as the ones they're replacing.
This change is likely to affect a lot of people, since products like Excel and Outlook are so widely used. It'll be worth keeping an eye on how this plays out, especially for anyone who relies on Copilot for their work.
The fact that Microsoft is already using their own models for tens of thousands of queries per week suggests that they're pretty confident in their ability to replace the external models. But it's still a big change, and it'll be interesting to see how it affects the overall performance of their products.
It's also worth thinking about what this means for the future of AI development. If Microsoft is able to successfully replace external models with their own, it could have big implications for the industry as a whole.
