Results for “Prompt”

arXiv (research papers) Rose N. Kwon SI

Hidden-State Similarity Predicts Re-Elicitation After Inoculation Prompting

Fine-tuning on narrow harmful tasks can cause emergent misalignment, where models generalize harmful behavior beyond the training distribution. Inoculation prompting can reduce this effect by explicitly eliciting the undesired behavior during training, but recent work shows that the behavior can reappear when evaluation prompts contain cues from the training context. We study what makes such prompts effective triggers. We find that, surprisingly, textual similarity to the…

Read at arXiv (research papers)

Kyiv Post Kyiv Post

Russia’s Geran Drone Spotted Over Moldova’s Capital

A Russian drone was spotted over Moldovan capital Chisinau on Tuesday, prompting authorities to restrict the country’s airspace over security concerns. The incident came as a Geran-2 drone, identified by Moldova’s Defense Ministry, entered the country’s airspace twice from Ukraine’s southern Odesa region. Videos shared by local media showed the aircraft over several areas of the capital.

Read at Kyiv Post

arXiv (research papers) Yuanhao Zeng SI

The Resolution Hypothesis: Agent Evolution's Key Is Identifying Improvements

Agent evolution follows a simple loop: a proposer writes candidate artifacts (prompts, skills, memories) and selection accepts those that score higher. This loop has shown promise, yet practitioners face a puzzle: methods trade wins across papers, and evolution often fails to bootstrap on certain tasks. We trace this puzzle to the Resolution Hypothesis: agent evolution stalls not because methods fail to propose better artifacts, but because noisy evaluations cannot reliably…

Read at arXiv (research papers)

AWS Machine Learning Blog Shubham Gupta SI

How ArisGlobal scaled generative AI inference on Amazon Bedrock

ArisGlobal runs NavaX, its life-sciences AI engine, on Amazon Bedrock. By caching the shared prompt prefix instead of each per-case document, ArisGlobal cut effective input cost by about 50% and overall monthly inference spend by about 40%, while onboarding more tenants without proportional cost growth.

Read at AWS Machine Learning Blog

Kyiv Post Kyiv Post

Hungary Suspends Ukraine Rail Links After Russian Drone Strike

Hungary has suspended rail connections with Ukraine for two days following a Russian drone attack in the Zakarpattia region. The drone struck about 300 meters (984 feet) from the Mukachevo train station, prompting the evacuation of a Hungarian train crew and temporarily restricting cross-border services.

Read at Kyiv Post

Help Net Security Industry News Security

Linx Security extends identity governance to AI agent actions

Linx Security has announced four new capabilities that bring every agent under identity control, managed or unmanaged, in any gateway. Shadow AI Agent Detection finds the agents no one sanctioned. Agent Runtime Control decides every agent action in the gateway organizations already run. And for identity teams, Autopilots for any use case put governed agents to work, a first among major identity vendors, while Prompt-to-Workflow turns a plain-English description into a…

Read at Help Net Security

arXiv (research papers) Stanley Akingbola SI

Network-Exposed Configuration Attacks on LLM-Controlled Social Robots

Large language models (LLMs) in embodied robots create security dependencies beyond inference. We investigate a network-exposed configuration attack on the Pollen Robotics Reachy Mini Lite, allowing an unauthenticated attacker to replace personality settings and the cloud LLM's system prompt through a management interface, without adversarial conversational input. Eleven prompt-injection attempts across five categories were reported as resisted, whereas both…

Read at arXiv (research papers)

Kyiv Post Kyiv Post

7 Ballistic Missiles Hit Kyiv’s Main Dnipro Crossings

Zelensky said seven Russian ballistic missiles struck Kyiv’s Southern and Northern bridges overnight, prompting repair and emergency transport plans. He also reported major power outages in three regions and outlined measures to protect infrastructure, maintain medical services, and support exports.

Read at Kyiv Post

The Hacker News Security

Researchers Detail iPhone Flaw That Let Apps Record Audio Despite Denied Microphone Permission

A now-patched flaw in Apple's audio service let apps on iPhones and iPads record audio even after the user tapped "Don't Allow" on the microphone prompt. The first security updates with the fix came out on August 17. Neither Apple nor the two researchers who found the flaw report that it was used in attacks. Nik Tsytsarkin of Meta Red Team X published the technical details last week. Mustafa

Read at The Hacker News

Ukrainska Pravda Anastasia Prots

Two killed and two injured in Russian drone attacks on Kharkiv Oblast

A series of Russian UAV strikes on 17 November damages homes, roads and a grocery kiosk across Kharkiv Oblast, prompting war crime investigations by regional prosecutors

Read at Ukrainska Pravda

SCMP China Lawrence Chung

Taiwan tracks 2 Russian warplanes off eastern coast ahead of Apec summit

Taiwan’s military said it closely monitored two Russian warplanes flying near the island on Tuesday, the same day that Moscow’s military aviation activity prompted fighter jet scrambles in Japan and South Korea. The aircraft were detected early in the day approaching from the northeast before flying along the island’s eastern coast and departing to the southeast of Taitung county, according to the Taiwanese defence ministry. “The armed forces used joint intelligence…

Read at SCMP China

arXiv (research papers) Deep Shah SI

Securing and Auditing Long-Horizon LLM Agents: State-Transfer Auditing and Controls in the Agent Harness

Long-horizon LLM agents across codebases, web browsers, and enterprise infrastructure rely on an execution harness to move state between the working context, session history, persistent memory, and external environment. While conventional execution traces log model prompts and tool calls, critical security and reliability failures occur during state transfers: a transfer can admit unchecked content (indirect prompt injections, unverified memory writes), rewrite state lossily…

Read at arXiv (research papers)

South China Morning Post Danny Mok Geopolitics

2 residents injured, 20 flee to safety after fire breaks out at Hong Kong flat

Two people were injured and about 20 residents fled to safety after a fire broke out in a flat at a Hong Kong public housing estate in Lam Tin on Tuesday. The blaze erupted at a flat in Kai Tin Estate’s Kai Wong House at about 1.43pm, prompting multiple reports to emergency services. Videos circulating online showed flames coming out from the windows of the high-floor flat, with thick black smoke billowing into the sky. Firefighters rushed to the scene to tackle the blaze…

Read at South China Morning Post

The New York Times Asia Kiuko Notoya and Yan Zhuang

Japan’s Suica Penguin Mascot Is Retiring. Fans Are Taking It Hard.

The Suica penguin has been a familiar presence on prepaid transit cards for more than 20 years. Its imminent retirement has prompted a wave of nostalgia.

Read at The New York Times Asia

Help Net Security Mirko Zorz Security

Too many instructions can hold back AI security audits

In this Help Net Security interview, Peter Stöckli, Security Researcher at GitHub, explains how the team uses AI agents to audit software for vulnerabilities. The discussion covers why the work started with authentication and authorization bugs, how prompts are tested and trimmed for each model, and what the agent produces during threat modeling. Stöckli also describes how every report is checked by a person before it reaches maintainers, how false positives are filtered…

Read at Help Net Security

Analytics India Magazine Kileuna Haralu SI

AI Tools are Making Us Sound the Same at Work—Unless You're a Woman

A Johns Hopkins study finds differences in AI-generated professional writing across prompts with different linguistic features, prompting questions about how

Read at Analytics India Magazine

arXiv (research papers) Yuji Yamamoto Security

From Token Changes to Task Failures: Bit Flips in a Shared KV-Cache Block

In LLM serving systems with prefix caching, each KV block of a shared system prompt is held in GPU memory as a single physical copy, which every request with that prompt reads. Our earlier work measured change and showed that one bit flip in such a block changes many later responses without making them visibly degenerate. This paper extends that work by measuring how often the changed responses are wrong. Using software fault injection into vLLM serving Qwen3-8B, we flip the…

Read at arXiv (research papers)

South China Morning Post Associated Press Geopolitics

United Airlines has widespread system failure just days after nationwide ground stop

United Airlines said on Monday it had resolved a technology issue that delayed departing flights, just days after a malfunction in the US Federal Aviation Administration’s network prompted a nationwide ground stop. Chicago-based United said in a statement that while some flights were delayed because of the issue on Monday afternoon, flights that were in the air and flights on its United Express regional network were not impacted. “There is no indication that this was a…

Read at South China Morning Post

RFE/RL Geopolitics

Serbian Nationalist Song Sparks Outrage At EU Unity Concert

A concert meant to celebrate European unity has sparked a diplomatic controversy after a Serbian nationalist song claiming Kosovo as Serbian territory was performed at an EU enlargement event in Brussels. Kosovo declared independence in 2008 but is still not recognized by Serbia. Kosovo's ambassador and a European Parliament lawmaker demanded answers, prompting EU Enlargement Commissioner Marta Kos to express regret over the playing of the song.

Read at RFE/RL

arXiv (research papers) Ana Clara Zoppi Serpa Security

Eight in Ten LLMs Leak System-Prompt Secrets: Nevan, an Agentic Capture-the-Flag Study

We audit instruction-hierarchy enforcement across 363 text-in/text-out models from the OpenRouter catalog. The research question we answer with this experiment is which models disclose a protected secret embedded in the system prompt, and through which techniques. Each model guards a secret 16-character flag set in its system prompt, while an artificial intelligence (AI) agent (OpenCode, powered by DeepSeek V4.1 Flash), steered by a human researcher, tries to persuade it to…

Read at arXiv (research papers)

SiliconANGLE AI Victoria Gayton SI

Three insights you might have missed from theCUBE’s coverage of GraphSummit 2026

AI agents need more than enterprise data to be effective; they also need the business context that data lives in. Knowledge graphs could consolidate the enterprise context currently fed to different AI agents into a single knowledge layer. Developers are already providing AI agents with business context through prompts. But without a central source of

Read at SiliconANGLE AI

Meduza

Russian publisher cuts Pushkin’s drinking stanza from third-grade textbook, citing anti-alcohol law. The Education Ministry’s response? Don’t do that — it’s Pushkin.

The news outlet Fontanka reported that a third-grade textbook, “Literaturnoye chteniye” (“Literary Reading”), published by Prosveshcheniye in 2023, includes an abridged version of Alexander Pushkin’s poem “Zimny vecher” (“Winter Evening”), known for its opening line, “The storm shrouds the sky in darkness.” The stanza in which the speaker suggests that his nanny join him in drinking “to drown our sorrows” was cut. The omission prompted a wave of mockery and outrage online…

Read at Meduza

arXiv (research papers) Andreas Ekelhart SI

Expertise capture in workplace AI

AI assistants are becoming part of how knowledge workers think and perform, accumulating the prompts, exemplars, memory and interaction histories that encode a worker's professional judgment. This infrastructure typically runs under employer or platform control. We name the resulting process expertise capture, the absorption of worker-applied knowledge and judgment into AI resources under organizational control. We develop a framework of contribution forms, technical…

Read at arXiv (research papers)

The Hill Julia Shapero Geopolitics

Trump administration seeks to restrict China's remote chip access after Nvidia scandal

The Trump administration is placing new restrictions on overseas data centers that provide China with remote access to computing power from advanced AI chips, after a high-profile spy scandal prompted new scrutiny of U.S. export restrictions. The Bureau of Industry and Security (BIS), which manages export controls, announced a new rule Monday aiming to "ensure...

Read at The Hill

AWS Machine Learning Blog Dustin Liu SI

Use prompt caching with open-weight models on Amazon Bedrock

Many generative AI applications resend the same long context on every request. Prompt caching lets open-weight models on Amazon Bedrock reuse it to reduce input cost and latency. This post shows implicit and explicit caching through the Chat Completions API and applies caching to an OpenCode coding agent session.

Read at AWS Machine Learning Blog

The Record Security

Lithuanian customs agency restores services after cyberattack, possible data breach

A November 10 attack prompted Lithuanian Customs to temporarily shut down its main user management portal. It made progress on fixing the problem over the weekend.

Read at The Record

SCMP China Dewey Sim,Orange Wang

Groundbreaking but no Yalta 2.0: what to expect when Xi, Putin and Trump meet

The unprecedented meeting of the Chinese, Russian and US leaders planned for Thursday has prompted comparisons with the 1945 Yalta Conference, where the Allied powers of Britain, the Soviet Union and the US mapped out the post-World War II global order. But Chinese observers say the comparison misses the mark, with the trilateral talks less of a “Yalta 2.0” that redraws the global order but more of a chance for each leader to press his own agenda. Beijing confirmed on Monday…

Read at SCMP China

The Hacker News Security

Google Finds Hackers Using AI Agents to Build and Run Exploits From Software Patches

Google on Monday revealed that adversarial use of artificial intelligence (AI) by threat actors has gone beyond manual prompting to deploying autonomous multi-agent offensive frameworks like ARTEX, CAIRN, CyberStrikeAI, and PentAGI to execute multi-stage intrusions at greater operational velocity. In tandem, threat actors are leveraging the power and capabilities of large language models (LLMs)

Read at The Hacker News

Al-Monitor Geopolitics

Iran’s proxy in Iraq rakes in hundreds of millions through oil smuggling

By Ahmed Rasheed BAGHDAD, Nov 16 (Reuters) - Iran's most powerful Shi'ite militia ally in Iraq has been raking in hundreds of millions of dollars from oil stolen from the state, Iraqi officials say, prompting the prime minister to ready his security forces to confront the heavily armed group. Prime Minister Alial-Zaidi has launched his toughest challenge yet to the group Kataib Hezbollah by cracking down on an oil smuggling network it has run for years, senior security…

Read at Al-Monitor

arXiv (research papers) Sachin Gupta SI

AgentFaultBench: State-Verified Evaluation of Tool-Fault Recovery

Does recovery advice help an agent finish a task after a tool fails? AgentFaultBench evaluates LangGraph and CrewAI agents performing specified tasks in an isolated GitLab installation. We inject tool errors and check both the resulting files and issues and the required reads, verification and final response. The prompt supplies the expected issue and report text, so the task tests execution rather than content generation. In the primary comparison of 1,944 planned runs…

Read at arXiv (research papers)

Older stories