UN negotiators in Geneva end a week of talks without agreeing to binding rules for lethal autonomous weapons. They settle on two years of further discussion toward a possible instrument by 2028, which could be legally binding or voluntary. A group led by Austria, Belgium, Brazil, Ireland, Mexico, New Zealand and Panama had pushed to begin treaty negotiations; Russia resisted stronger steps, while the United States had proposed five years of discussions on optional guidelines. Under the consensus process of the Convention on Certain Conventional Weapons, any one of its 128 member countries can block an agreement. The outcome falls short of calls from UN Secretary-General AntĂłnio Guterres, many governments and humanitarian groups, including the International Committee of the Red Cross, for binding rules.
Companies face growing security concerns as AI agents connect to internal systems. Forgotten API keys, developer sandboxes and unapproved integrations can leave agents with access to data, sometimes through credentials that conventional security tools may mistake for normal application activity. The article cites a breach involving third-party AI tool Context.ai that led to an employee’s Google Workspace account being taken over and access to a Vercel account. Separately, security experts warn that AI helps attackers find and exploit software flaws faster and at lower cost, while AI-generated code may go unreviewed. Red-team testing can probe risks such as direct and indirect prompt injection and jailbreaking, including attempts to make chatbots disclose private or other users’ information.
In healthcare, the American Medical Association calls for laws that clarify responsibility when AI tools cause harm. Its framework, released Oct. 29, says doctors should be able to refuse tools when their safety or clinical value is uncertain. The AMA also wants technology companies to carry liability insurance and clearly disclose product limits and performance; contracts and patient consent forms should not erase developers’ responsibility. Separately, research into brain cancer treatments faces major hurdles: glioblastoma survival often increases by only four to six months with standard treatment, and the blood-brain barrier limits drug delivery. Phase III trials of checkpoint inhibitors including nivolumab, pembrolizumab and ipilimumab did not significantly improve overall survival among unselected glioblastoma patients. The article also discusses oncolytic viruses and cell-based approaches.
AI may make it cheaper to launch a company, but not to win customers, argues Entrée Capital’s Yoni Osherov. Wix acquired Base44 in June 2025 for about $80 million in initial consideration, with possible earn-out payments; founder Maor Shlomo said Wix could bring scale and distribution while allowing the product to move quickly. Carta data shows solo founders accounted for 36.3% of new startups in its US sample in the first half of 2025, up from 23.7% in 2019. Osherov advises founders to raise money when it addresses a demonstrated constraint, such as market access, reliability or serving larger customers. The article also reports $515.8 billion in US venture deal value through the third quarter of 2026, while cautioning that this figure does not indicate what an individual company needs.
Meta’s Muse personal AI agent attracted more than 10 million users in the US and Canada during its first six weeks, with shopping its most common use. The article says familiar brands can come to mind first for consumers, while AI answers depend more directly on available information. Generative engine optimization (GEO) seeks to improve a brand’s presence in AI answers; the piece raises the possibility of shaping the questions people ask but does not explain how. In South Korea, authorities plan a 10-second phone alert on exam morning to help detect hidden devices before the college entrance test. Proctors will inspect glasses and seal shirt buttons with stickers, while 8,400 additional supervisors help monitor more than 500,000 students.
Director Luca Guadagnino discusses Artificial, his film about OpenAI, starring Andrew Garfield as Sam Altman and Monica Barbaro as Mira Murati. Neon plans a limited release on Christmas Day following screenings at the New York Film Festival and BFI London Film Festival. Amazon MGM Studios left the project, reportedly amid concerns about OpenAI. Guadagnino says the film’s facts were legally checked and that he faced no pushback from Altman or OpenAI. Altman has not seen the film; he said its trailer was not aiming for accuracy and disputed its depiction of a gun-filled bunker at his wine-country home.
Key Takeaways
- UN talks produce no binding treaty on lethal autonomous weapons; countries agree to continue discussions toward a possible instrument by 2028.
- Russia opposed stronger steps, and the consensus process means any of the 128 countries party to the weapons convention can block an agreement.
- Unmanaged AI agents can expose company data through forgotten keys and integrations; a Context.ai breach led to a Google Workspace account takeover and access to Vercel.
- Security experts warn AI accelerates vulnerability discovery and exploitation, while AI-written code may go unreviewed; red teams test prompt injection and jailbreak risks.
- The AMA wants laws clarifying liability for medical AI errors, developer insurance and disclosures, and the right for doctors to refuse uncertain or unsafe tools.
- Wix acquired Base44 in June 2025 for about $80 million in initial consideration, with possible earn-out payments; its founder cited Wix’s scale and distribution.
- Entrée Capital’s Yoni Osherov says founders should raise capital to address proven constraints, not simply because funding is available.
- Meta’s Muse drew more than 10 million users in the US and Canada in its first six weeks, with shopping its most common use; brands are exploring GEO to appear in AI answers.
- South Korea plans phone alerts and tighter device checks to deter AI-assisted cheating during an exam taken by more than 500,000 students.
- Amazon MGM Studios exited Luca Guadagnino’s Artificial, reportedly amid concerns about OpenAI; Neon plans a limited Christmas Day release.
UN talks on autonomous weapons end without a binding treaty
Negotiators at a United Nations conference in Geneva failed to agree on a plan to negotiate a legally binding treaty on autonomous weapons. After a week of talks, countries agreed only to discuss a possible instrument by consensus, which may not create legal duties. Russia objected to stronger steps, while many countries and humanitarian groups wanted binding rules. A group led by Austria, Belgium, Brazil, Ireland, Mexico, New Zealand and Panama pushed for treaty negotiations.
UN talks set two years for discussion of rules on killer robots
Diplomats at the United Nations in Geneva agreed to spend two years considering rules for lethal autonomous weapons, which are often called killer robots. Any rules could be legally binding or optional, falling short of the International Committee of the Red Cross’s call for binding rules within a year. Humanitarian groups fear that AI weapons could make decisions to kill without human control. Russia delayed the talks, and the United States proposed five years of discussion on optional guidelines before negotiators reached their compromise.
UN nations delay binding rules for AI killer robots until 2028
Countries at a weeklong United Nations conference in Geneva agreed to continue work on a possible deal for lethal autonomous weapons by 2028, but they did not commit to binding rules. UN Secretary-General AntĂłnio Guterres and many governments and activists had called for stronger action. Brazil, Austria, Belgium, Ireland, Mexico, New Zealand and Panama had proposed starting treaty negotiations, but their effort fell short. The talks use consensus under the Convention on Certain Conventional Weapons, ratified by 128 countries, so any one country can block an agreement.
Untracked AI agents can create hidden paths to company data
AI agents connected to company systems can gain access through forgotten API keys, developer sandboxes and unapproved software integrations. If an agent has broader permissions than needed and is left unmanaged, an attacker who finds it could use those valid credentials to reach company data. Traditional security tools may not flag activity that looks like normal use of an application. AI integrations can connect to several systems at once, and projects may leave behind data copies, indexes and logs outside governance controls. The article also describes a breach involving third-party AI tool Context.ai, which led to the takeover of an employee’s Google Workspace account and access to a Vercel account.
AMA seeks laws clarifying responsibility for medical AI errors
The American Medical Association is calling for laws that clarify who is responsible when AI tools make mistakes in healthcare. AMA President Willie Underwood III said AI can support clinical decisions, but patients should know who is responsible for their care. The AMA also wants technology companies to carry liability insurance and provide clear information about product limits and performance. It says provider contracts, patient disclosures and consent forms should not remove developers’ responsibility. On Oct. 29, the AMA released a framework saying doctors should be able to refuse AI tools without penalty when their clinical value or safety is uncertain.
Nanotechnology and AI offer new paths to treat deadly brain cancers
Brain and central nervous system cancers caused 321,731 new cases and 248,500 deaths worldwide in 2020. Glioblastoma is an aggressive brain tumor, and standard treatment with surgery, radiation and chemotherapy often extends survival by only four to six months. The blood-brain barrier blocks many drugs, while the tumor environment also weakens immune responses. Researchers have tested treatments such as nivolumab, pembrolizumab and ipilimumab, but phase III trials of checkpoint inhibitors did not significantly improve overall survival in unselected glioblastoma patients. The article also discusses oncolytic viruses and cell-based treatments, but the supplied text ends before giving further details.
AI tools are making cyberattacks easier and software failures harder to stop
Artificial intelligence is helping attackers find weaknesses in websites and software more quickly and at lower cost. Security experts say AI tools can let less-skilled hackers do work that once required a team, while vulnerabilities may now be exploited within days or hours. The growing number of exposed flaws has been called the “vulnpocalypse.” AI can also write software, but companies may use smaller teams and code that no person has reviewed. Black Hills Information Security CEO Strand says AI agents can automate routine attack work and help human operators move faster.
AI makes it cheaper to start a company, not to win customers
Yoni Osherov of Entrée Capital argues that AI lowers the cost of building a company but does not remove the challenge of reaching customers and growing. Wix acquired Base44 in June 2025 for about $80 million in initial consideration, with possible earn-out payments; founder Maor Shlomo said Wix could provide greater scale and distribution while keeping the product moving quickly. Carta data shows solo founders made up 36.3% of new startups in its US sample in the first half of 2025, up from 23.7% in 2019. PitchBook reported $515.8 billion in US venture deal value through the third quarter of 2026, but that total does not show how much funding a particular company needs. Osherov says founders should raise money when it can address a proven limit, such as market access, reliability or serving larger customers, rather than raising simply because they can.
Ten security risks companies should test in AI applications
AI red teaming tests how attackers could manipulate large language model applications into unsafe or unauthorized actions. Unlike traditional vulnerability testing, it checks system prompts, data retrieval, model behavior and connected tools. For example, attackers may try to make an enterprise chatbot reveal private documents or information belonging to another user. The article outlines 10 risks organizations should test, including direct prompt injection, indirect prompt injection and jailbreaking. The provided text explains these first risks but ends before listing the rest.
Brands can shape the questions people ask AI search tools
Personal AI agents such as Meta’s Muse can use information from sources like email inboxes to guide their responses. More than 10 million people in the US and Canada started using Muse in its first six weeks, and shopping was its most common use. The article says people often choose familiar brands because those brands come to mind first, while AI tools rely more literally on available information. Generative engine optimization, or GEO, aims to make a brand appear in AI answers. The article suggests brands could go further by influencing the questions people ask, though the supplied text does not explain how.
South Korea Tightens Exam Checks to Stop AI Cheating
South Korea plans to send a 10-second alert to nearly every phone on Thursday morning to help detect hidden devices before its college entrance exam. Proctors will inspect students’ glasses and seal shirt buttons with stickers to block cameras. An additional 8,400 supervisors will help monitor more than 500,000 students. The article also notes that England’s Office for Students recently said AI has made it ...
Luca Guadagnino Discusses Sam Altman and AI in Film Artificial
Director Luca Guadagnino spoke with Deadline about his film Artificial, which stars Andrew Garfield as Sam Altman and Monica Barbaro as Mira Murati. Neon plans a limited release on Christmas Day, after screenings at the New York Film Festival and BFI London Film Festival. Amazon MGM Studios left the project, reportedly amid concerns about OpenAI, while Guadagnino said the film’s facts were checked legally and that he faced no pushback from Altman or OpenAI. Altman has not seen the movie, but said its trailer was not aiming for accuracy; he also disputed the trailer’s depiction of a bunker filled with guns at his wine country home.
Sources
- International talks on controlling use of autonomous weapons collapse
- Killer robots: U.N. talks fall short of hopes for legally binding rules
- U.N. talks fail to reach firm commitment to rein in AI killer robots
- Shadow AI: How agents bypass data governance programs
- AMA calls for laws to protect physicians from AI-related lawsuits
- Nanotechnology and Artificial Intelligence Advance Immunotherapy Against Deadly Brain Cancers
- Everything Is Falling Apart, and AI's Fingerprints Are All Over It
- AI lowered the cost of starting a company, but not the cost of winning
- AI Red Teaming for LLM Applications: 10 Security Risks Organisations Should Test
- To win the AI search wars, stop chasing the answers and start shaping the questions
- As AI cheating evolves, exam halls are taping shirt buttons and glasses to block cameras
- 'Artificial's Luca Guadagnino On Sam Altman's Arsenal; AI As Doomsday
Comments
Please log in to post a comment.