Recent research has demonstrated that state-of-the-art LLMs and defenses remain susceptible to multi-turn jailbreak attacks. These attacks require only closed-box model access and are often easy to perform manually, posing a significant threat to the safe and secure deployment of LLM-based systems. We study the effectiveness of the Crescendo multi-turn jailbreak at the level of intermediate model representations and find that safety-aligned LMs often represent Crescendo responses as more benign than harmful, especially as the number of conversation turns increases. Our analysis indicates that at each turn, Crescendo prompts tend to keep model outputs in a "benign" region of representation space, effectively tricking the model into fulfilling harmful requests. Further, our results help explain why single-turn jailbreak defenses like circuit breakers are generally ineffective against multi-turn attacks, motivating the development of mitigations that address this generalization gap.

Large language models (LLMs) in the form of chatbots like ChatGPT and Llama are increasingly proposed as "silicon samples" for simulating human opinions. This study examines this notion, arguing that LLMs may misrepresent population-level opinions. We identify two fundamental challenges: a failure in structural consistency, where response accuracy doesn't hold across demographic aggregation levels, and homogenization, an underrepresentation of minority opinions. To investigate these, we prompted ChatGPT (GPT-4) and Meta's Llama 3.1 series (8B, 70B, 405B) with questions on abortion and unauthorized immigration from the American National Election Studies (ANES) 2020. Our findings reveal significant structural inconsistencies and severe homogenization in LLM responses compared to human data. We propose an "accuracy-optimization hypothesis," suggesting homogenization stems from prioritizing modal responses. These issues challenge the validity of using LLMs, especially chatbots AI, as direct substitutes for human survey data, potentially reinforcing stereotypes and misinforming policy.

Deploying robots in human-centric settings like households requires balancing robot autonomy with humans' sense of agency [1, 2, 3, 4, 5, 6]. Full teleoperation offers users fine-grained control but imposes a high cognitive load, whereas fully autonomous robots act independently but often misalign their actions with nuanced human needs. Assistive teleoperation -- a paradigm in which both the human and the robot share control [7, 8, 9, 10] -- has thus emerged as an ideal middle ground. By keeping the user in control of high-level decisions while delegating low-level actions to the autonomous robot, this approach both preserves user agency and enhances overall system performance. As such, assistive teleoperation is becoming a desirable paradigm for robots to serve as reliable partners in human-centric environments, such as assisting individuals with motor impairments [11, 12]. While promising, assistive teleoperation in everyday environments remains challenging. A longstanding challenge in assistive teleoperation is to infer human intents from user control inputs and assist users with correct actions [8]. This challenge is amplified in real-world settings, where robots must go beyond closed-set intent prediction [13, 14] to handle diverse, open-ended user goals across different contexts and scenes. As a result, a key capability the robot should possess is to interpret user control inputs within the visual context and infer intent through commonsense reasoning.

Breakthroughs, discoveries, and DIY tips sent every weekday. A US astronaut aboard the International Space Station (ISS) recently caught a glimpse of one of Earth's least understood atmospheric phenomena. While orbiting in the early hours of July 3, Nichole "Vapor" Ayers snapped a photo of a transient luminous event, as she passed over North America. Better known as a sprite, these atmospheric events are common after a lightning strike. Wow," Ayers posted to social media later that day along with the stunning picture.

A senior United States envoy has praised the Lebanese government's response to a US proposal aimed at disarming Hezbollah amid Israel's continued military presence in the country. Thomas Barrack, an adviser to US President Donald Trump who serves as Washington's ambassador to Turkiye and special envoy for Syria, returned to Beirut on Monday after delivering the US proposal during a June 19 visit. The plan called for the Shia Lebanese group Hezbollah to fully disarm within four months in exchange for a halt to Israeli air strikes and the full withdrawal of Israel's military from the five positions it continues to occupy in southern Lebanon. "What the government gave us was something spectacular in a very short period of time," Barrack told reporters on Monday after meeting Lebanese President Joseph Aoun. "I'm unbelievably satisfied with the response." While Barrack confirmed that he had received a seven-page reply from the Lebanese side, he offered no details on its contents.

A terrifying new scam is sweeping across the US, catching people off guard with fake threats of arrest and demands for payment. The FBI issued a warning about criminals impersonating federal officials in phone calls targeting potential victims. In some cases, callers falsely claim a victim's Social Security number is frozen or that a loved one must pay for GPS monitoring to be released from jail, creating a false sense of urgency. The scam relies on phone number spoofing technology, which falsifies caller ID to make it appear as though the call is coming from trusted government agencies. Federal agents say scammers are demanding thousands of dollars in payment via prepaid gift cards, wire transfers, or cryptocurrency ATMs.

Drew Kesse's daughter, Jennifer, vanished from her Central Florida apartment in 2006, with FDLE recently announcing the cold case has been reopened with new evidence. Nearly two decades after a Florida woman vanished from her apartment building without a trace, authorities have announced a new break that could breathe new life into a formerly cold case. Jennifer Kesse, 24, vanished from her Orlando condo complex after leaving for work on the morning of Jan. 24, 2006, stumping both state and federal investigators as authorities raced to catch her abductor. "About an hour and a half into the workday, I received a call from her work," Drew Kesse, Jennifer's father, told Fox News Digital. "And they said, 'Hey, Jennifer had a meeting this morning, it's not like her to show up. Do you know where she is?'" Jennifer Kesse smiles in an undated family photo.

United Nations Secretary-General Antonio Guterres has condemned a Russian drone and missile attack against Ukraine this week that has been described as the largest such assault in the three-year war. In a statement on Saturday, Guterres's spokesperson said the Russian strikes "disrupted the power supply to the Zaporizhzhia Nuclear Power Plant, once again underlining the ongoing risks to nuclear safety". "The secretary-general is alarmed by this dangerous escalation and the growing number of civilian casualties," the statement read. Ukrainian officials said Moscow fired more than 500 drones and 11 missiles at the capital Kyiv overnight into Friday in an attack that killed one person, injured at least 23 others and damaged buildings across the city. The sounds of air raid sirens, kamikaze drones and booming detonations reverberated until dawn.

Thanks to a team at the University of California, Davis, theres a new brain-computer interface (BCI) system thats opening up real-time, natural conversation for people who cant speak. VOICE BREAKTHROUGH: When someone loses the ability to speak because of a neurological condition like ALS, the impact goes far beyond words. Now, thanks to a team at the University of California, Davis, there's a new brain-computer interface (BCI) system that's opening up real-time, natural conversation for people who can't speak. Instead, it translates the brain signals that would normally control the muscles used for speech, allowing users to "talk" and even "sing" through a computer, almost instantly. JOBS ON THE LINE: If you've ordered food on Uber Eats recently, you may have seen a delivery robot instead of a human driver.

Israel has carried out four drone attacks on towns in southern Lebanon, resulting in a death and several injured, in the latest wave of near-daily Israeli violations of the November ceasefire between Israel and the Lebanese group Hezbollah. An "Israeli enemy drone attack on a vehicle" in the Saf al-Hawa area in the city of Bint Jbeil "killed one person and wounded two others", Lebanon's Ministry of Public Health said in a statement on Saturday carried by the official National News Agency (NNA), noting the toll was expected to rise. A second attack in the Bint Jbeil area followed. Earlier Saturday, the ministry also reported that a separate Israeli drone attack wounded one person in Shebaa, with the NNA saying that raid hit a house. Shebaa is located across two steep, rocky mountainsides that straddle Lebanon's borders with Syria and the Israeli-occupied Golan Heights.