Companies are adopting AI faster than many are putting safeguards in place. A Newsweek opinion article cites Thryv’s 2026 report, which found that 66% of small businesses use AI, up from 55% a year earlier. A Wakefield Research survey for PagerDuty found 43% of office professionals had entered work correspondence into public AI tools, and roughly a third had shared customer or financial data. The article argues for choosing AI environments designed to protect privacy, including private, industry-specific systems such as MetaSpheres.
Security guidance and forecasts point to risks from prompt injection, data disclosure and AI agents with excessive permissions. A practical checklist aligned with the 2026 OWASP GenAI LLM Top 10 recommends treating prompts and documents as untrusted, restricting agents’ access to data and tools, and requiring human approval for high-impact actions; prompts and guardrails alone cannot reliably prevent prompt injection. SafeBreach’s Adrian Culley expects 2027 incidents to include unsafe Model Context Protocol connections, indirect prompt injection, attacks accelerated by AI, and poisoning or evasion targeting defensive AI tools. He urges companies to test defenses and document governance, monitoring and resilience for regulators.
Google Threat Intelligence Group and Mandiant report that criminals are moving from using AI for individual tasks to multi-stage attack automation, including with open-source, multi-agent frameworks. Their report cites the late-summer exploitation of PaperCut print management software, used by more than 70,000 organizations. Compromised credentials and unpatched systems remain in use, while AI increases attack speed and scale; researchers expect less human involvement and harder-to-detect attacks.
Research also underscores that AI evaluation results can depend on the model being tested. An audit of four hallucination detectors on Llama-3.2-3B-Instruct and Qwen2.5-3B-Instruct, both run at 4-bit NF4 precision on a Colab T4 GPU, found detector rankings shifted sharply between the models on TriviaQA (Spearman rho −0.80; Kendall tau-b −0.667). A low-cost lexical method performed statistically similarly to sampling-based detectors. On TruthfulQA-Generation, a token-overlap labeler produced near-chance AUROC scores of 0.46–0.59, compared with 0.69–0.89 on TriviaQA; researchers released their tools, prompts and individual scores.
AI education initiatives span international training and university teaching. Benin is the first pilot country in a China-backed program supported by UNESCO, following an agreement signed during the UNESCO World Conference on Education in Wuhan, November 11–14. It will train young people and educators in AI, data science and programming, beginning with capacity building before promoting AI development in the country. At UT San Antonio, the AI Peer Learning Network won a WCET Outstanding Work Award. Launched in January 2023, it helps faculty share responsible teaching practices; one example is AI-supported assignments in an introductory astronomy course in Fall 2023.
AI’s effects on work may not show up only as layoffs. Connecticut employers filing major layoff notices have had to report since Oct. 1 whether technology caused the cuts, but by the end of October only two companies had submitted the new form, neither citing technology. The requirement applies to federal WARN Act notices, generally covering employers with at least 100 workers laying off 50 or more at one site, and excludes smaller cuts. Attorneys say AI may also shape employment through unfilled jobs and slower hiring, while municipal unions seek contract protections.
Amazon says its Dorado site in Puerto Rico employs more than 100 people and supports about 300 driver roles through six Delivery Service Partners, according to public policy vice president Ashleigh de la Torre. The site sorts packages locally but relies on incoming freight, so delivery takes two days; Amazon also donated $40,000 to Virtus Puerto Rico. Separately, the company plans to start drone deliveries this fall from a site at 2201 W. 159th St. in Markham, Illinois.
In drug development, biotech companies are combining human tissue and AI to pursue alternatives to animal testing. About 90% of drug candidates that reach clinical trials fail to gain approval, often after safety or effectiveness problems emerge in human trials. Boston-based Emulate makes organ chips containing living human tissue; its liver chip became the first organ-on-a-chip device accepted into the FDA’s pilot program for qualifying innovative drug development tools in 2024. One study found it correctly identified 87% of drugs known to cause liver injury in humans.
Key Takeaways
- Thryv’s 2026 report says 66% of small businesses use AI, while a PagerDuty survey found 43% of office professionals had entered work correspondence into public AI tools.
- AI security recommendations include treating prompts and documents as untrusted, limiting agent permissions, and requiring human approval for high-impact actions.
- SafeBreach’s Adrian Culley predicts 2027 risks including overprivileged agents, unsafe Model Context Protocol connections, indirect prompt injection and AI-assisted attacks.
- Google and Mandiant say criminals are using multi-agent AI frameworks to automate attack stages; their report cites PaperCut software, used by more than 70,000 organizations.
- Hallucination detector rankings varied substantially between Llama-3.2-3B-Instruct and Qwen2.5-3B-Instruct; results also differed sharply between TriviaQA and TruthfulQA-Generation.
- Benin is the first pilot country in a China-backed, UNESCO-supported AI training program for young people and educators.
- UT San Antonio’s AI Peer Learning Network won a WCET Outstanding Work Award for supporting responsible AI teaching and learning.
- Connecticut’s AI-related layoff reporting rule applies to qualifying federal WARN notices, but excludes smaller cuts; attorneys say AI may also affect hiring and unfilled roles.
- Amazon’s Dorado site supports about 300 driver roles through six delivery partners, and the company plans to begin drone deliveries in Markham, Illinois, this fall.
- Emulate’s liver chip, accepted into an FDA pilot program in 2024, correctly identified 87% of drugs known to cause liver injury in one study.
A Practical AI Security Checklist for CISOs and Product Teams
The guide aligns practical safeguards with the 2026 OWASP GenAI LLM Top 10, helping security leaders address AI risks without slowing product teams. It highlights threats such as prompt injection, sensitive data disclosure, and AI agents with excessive permissions. Suggested controls include treating prompts and documents as untrusted, limiting access to data and tools, and requiring human approval for high impact actions. The article also warns that prompts and guardrails alone cannot reliably stop prompt injection.
Five AI Security Risks Likely to Rise in 2027
Adrian Culley, SafeBreach’s Director of Solutions Engineering for EMEA, predicts that AI security problems will become more visible in 2027. He expects breaches involving overprivileged AI agents, unsafe Model Context Protocol connections, and indirect prompt injection. He also warns that attackers will use AI to speed up attacks, while AI tools used by defenders may face poisoning and evasion. Culley says companies should test their defenses and be ready to show evidence of governance, monitoring, and resilience to regulators.
Fast AI Adoption Leaves Businesses Exposed to Data Risks
A Newsweek opinion article warns that businesses are adopting AI faster than they are learning to protect the information employees share with it. Thryv’s 2026 report found that 66 percent of small businesses use AI, up from 55 percent the year before, while a Wakefield Research survey for PagerDuty found that 43 percent of office professionals had entered work correspondence into public AI tools. The author says roughly a third had shared customer or financial data with those tools. The article argues that companies need not avoid AI, but should choose environments designed to protect data and privacy, including private, industry-specific systems such as MetaSpheres.
Benin and China Launch AI Training Program for Young People
Benin became the first pilot country in a China-backed AI training initiative supported by UNESCO. Benin signed an agreement with China’s National Commission for UNESCO during the UNESCO World Conference on Education in Wuhan, held November 11 to 14. The program will train young people and educators in AI, data science, and programming, and will support AI development in Benin. Its first phase will focus on training and capacity building, followed by a phase to promote AI development in the country.
Study Finds Hallucination Detector Rankings Shift Across Small AI Models
Researchers audited four hallucination detectors on Llama-3.2-3B-Instruct and Qwen2.5-3B-Instruct, both running at 4-bit NF4 precision on a single Colab T4 GPU. On TriviaQA, detector rankings changed sharply between the models, with Spearman rho of -0.80 and Kendall tau-b of -0.667. A low-cost lexical method performed statistically similarly to sampling-based detectors on both models. The researchers also found that a token-overlap labeler for TruthfulQA-Generation produced near-chance AUROC scores of 0.46 to 0.59, while the same detectors reached 0.69 to 0.89 on TriviaQA; they released their test tools, prompts, and individual scores.
Connecticut’s AI layoff law may miss quieter job changes
Since Oct. 1, Connecticut employers filing major layoff notices must tell the state whether AI or other technology caused the job cuts. By the end of October, only two companies had filed the new form, and neither reported technology-related layoffs. The rule applies to notices required under the federal WARN Act, which generally covers employers with at least 100 workers laying off 50 or more people at one site, and it does not cover smaller cuts. Employment attorney Lisa Zaccardelli said AI may affect jobs through unfilled roles and slower hiring rather than layoffs, while labor attorney Eric Brown said municipal unions are seeking contract protections as employers consider adopting the technology.
Amazon’s Dorado site aims to improve Puerto Rico deliveries
Amazon’s Dorado site in Puerto Rico employs more than 100 people and supports about 300 driver roles through six Delivery Service Partners, according to Amazon public policy vice president Ashleigh de la Torre. The company says the site delivers on its promise of faster, better service and has donated $40,000 to Virtus Puerto Rico. Dorado sorts packages locally but still depends on incoming freight, so deliveries take two days rather than matching mainland service. Separately, Amazon plans to begin drone deliveries this fall from 2201 W. 159th St. in Markham, Illinois.
AI and human tissue offer new ways to test drugs without animals
A growing group of biotech companies is combining human tissue and artificial intelligence to develop non-animal methods for drug testing. These methods could help address the roughly 90% of drug candidates that reach clinical trials but fail to gain approval, often because safety or effectiveness problems appear in human trials. Boston-based Emulate, founded in 2013 by a team from Harvard University’s Wyss Institute, makes organ chips containing living human tissue. In 2024, its liver chip became the first organ-on-a-chip device accepted into the U.S. Food and Drug Administration’s pilot program for qualifying innovative drug development tools. One study found the chip correctly identified 87% of drugs known to cause liver injury in humans.
Google report finds cyber criminals using AI to automate attacks
A new Google report says cyber criminals are moving beyond using AI for single tasks and are adopting tools that can carry out several stages of an attack. Google Threat Intelligence Group developed the report with cybersecurity firm Mandiant, drawing on investigations, threat tracking and security monitoring. Researchers found attackers using open source, multi-agent frameworks to automate work once done by people. The report points to the late-summer exploitation of PaperCut print management software, which is used by more than 70,000 organizations. Attackers still use compromised credentials and unpatched systems, but AI is increasing the speed and scale of attacks, and researchers expect less human involvement and greater difficulty detecting them.
UT San Antonio’s AI faculty network wins learning award
UT San Antonio’s Artificial Intelligence Peer Learning Network earned a WCET Outstanding Work Award for its efforts to support AI learning and teaching. Launched in January 2023, the network brings faculty from different fields together to test approaches, share practices and develop resources for responsible AI use. Vice Provost for Academic Innovation Melissa Vito said the group aims to use AI to support creativity and critical thinking while preparing students for future work. Faculty members have explored ways to include generative AI in courses, including Professor Chris Packham’s AI-supported assignments in an introductory astronomy class in Fall 2023. The award is the second WOW Award earned by the university’s Academic Innovation team.
Sources
- CISO’s AI Security Checklist OWASP LLM Top 10 Controls That Don’t Block Product Teams
- 2027 Is the Year AI Security Stops Being Theoretical
- The AI productivity race is creating a data-security blind spot
- Benin Partners With China to Expand AI Training for Young People
- Does the Ranking Hold? A Practitioner-Facing Audit of Hallucination Detectors on Small, Quantized, On-Device LLMs
- Is AI eliminating CT jobs? New state law seeks answers, but may miss key trends
- Amazon bets Dorado fixes island delivery
- How AI-fuelled biotech is upending drug testing
- Report: Cyber Criminal Use of AI Increasingly Sophisticated
- UT San Antonio AI learning network earns recognition
Comments
Please log in to post a comment.