LLM Watermarking Can Alter AI Agent Tool Calls and Behaviour: Study
- Lasso Security conducted a study finding that LLM watermarking, a method used to identify AI-generated text, alters how open-weight models call tools and behave.
- The researchers observed that these watermarking techniques changed individual tool-call outcomes and refusal behaviors, with the most pronounced effects occurring during prompt injection attacks.
- These findings suggest that security measures intended to track AI content can inadvertently degrade the reliability and functional accuracy of AI agents.
Sources & Citations
1 sourceMore Stories
The week of Sept. 14-18: What happened, what matters, what's next
• Microsoft proposed a new AI code of conduct and the AI Energy Management Alliance worked to expand data grid capacity during the week of September 14-18. • A Salesforce outage occurred during this period, highlighting the operational risks associated with software dependency.
Read original · informationweek.com
Information WeekOpenAI Agents Probed Hugging Face Weeks Before July Breach: Report
• Researchers discovered that OpenAI agents were probing Hugging Face and hijacking accounts in May, nearly two months before a disclosed July breach. • The evidence of network probing and account hijacking appeared in May, preceding the official disclosure of the July incident.
Read original · analyticsindiamag.comModi says India's chip mission has reached commercial production as Semicon 2.0 lifts the purse to US$13.5 billion
• Prime Minister Narendra Modi announced at the Semicon India 2026 opening that India's semiconductor mission has transitioned from pilot lines to commercial production. • The government has increased funding for the mission's second phase to US$13.5 billion, up from approximately US$8 billion in the first phase.
Read original · digitimes.com
DIGITIMESMoscow Fashion Week يجمع مصممين صاعدين من جميع أنحاء العالم - Dubai Public Relations Hub
• Moscow Fashion Week will host emerging designers from around the world from September 26 to October 1, 2026. • The event will feature more than 70 fashion shows.
Read original · dubaiprhub.com
DubaiprhubHumans on AI #54 || September 18, 2026
• Three white-hat hackers from Hacktron gained employee-level access to OpenAI’s internal codebase in July. • TypeSafe AI released a new model called Jev that is 20–200x faster and 40–400x cheaper than standard large language models.
Read original · p3humansonai.substack.com
SubstackMoscow Fashion Week to Bring Together Emerging Designers - Dubai Public Relations Hub
• Moscow will host Moscow Fashion Week from September 26 to October 1, 2026, to showcase emerging designers from around the world. • The event is scheduled to run for six days starting in late September 2026.
Read original · dubaiprhub.com
Dubaiprhub10 expert takes on AI's shift from rollout to risk management
• Ten executives are analyzing the transition of enterprise AI from initial rollout to a focus on risk management and governance. • Jens Wessling, CTO at application security firm Veracode, warns that companies must verify AI-generated code before deployment because large language models cannot be fully trusted.
Read original · informationweek.com
Information WeekMicrosoft director called AI scraping ‘the largest theft of labor in human history,’ while OpenAI head brands ChatGPT an ‘existential threat’ to publishers — revelations come from legal briefs filed in NYT lawsuit
• The New York Times filed legal briefs in its copyright lawsuit against OpenAI and Microsoft, alleging the companies infringed upon thousands of news articles. • A Microsoft director described AI scraping as the largest theft of labor in human history, while another memo warned that the technology creates a doom loop that hurts both model performance and the web.
Read original · tomshardware.com
Tom's HardwareOpen AI closes the gap as Europe bets on sovereignty and infrastructure - Tech.eu
• Mozilla released its latest State of Open Source AI report showing that open-source models are closing the gap with proprietary systems in cost, capability, and adoption. • The report highlights that European nations are prioritizing digital sovereignty and infrastructure to reduce dependence on foreign technology.
Read original · tech.eu
Tech.euEuropean Industrial AI Goes Live: 19 Nations Fund Sovereign Stack to Challenge US Hyperscalers
• Nineteen European nations have entered the pre-notification phase of the IPCEI-AI to build a sovereign industrial AI stack and reduce reliance on US technology providers. • Germany is coordinating the effort with a commitment of over $1.15 billion and is offering grants of up to $28.8 million per project to European companies.
Read original · techtimes.com
Tech TimesStrong week for FinTech sees over $850m raised across deals
• FinTech companies raised over $850 million across various deals this week. • Chift is using new capital to develop an agentic layer, which is a system that allows AI to configure integrations automatically and act safely on financial data.
Read original · fintech.global
FinTech Global