Artificial Intelligence
The MIT-Licensed Frontier: Why GLM-5.2 Reshapes Enterprise AI Trade-Offs
Enterprise artificial intelligence strategy is shifting from model selection to control architecture selection.
As organizations move from experimental deployments toward production-grade agentic systems, the dominant constraints are no longer model performance alone. They are increasingly defined by control over weights, data residency, licensing structure, and operational governance boundaries.
The release of GLM-5.2 by Z.ai (Zhipu AI) reflects this shift. Based on publicly available technical documentation and reported benchmark evaluations, the model is positioned as a large-scale mixture-of-experts system targeting frontier-level capability in software engineering and multi-step reasoning tasks.
Its significance is not isolated performance. It is the combination of capability, deployment flexibility, and permissive licensing under a widely used open-source framework.
The first hurdle is the hardest in generative AI adoption – and businesses keep falling | IT Pro Despite rapid AI adoption, many businesses struggle with implementation, falling into “pilot purgatory” due to issues like skills gaps, legacy systems, and a lack of advanced use cases. While employees report individual productivity gains, companies are slow to achieve business-wide benefits, with a significant portion of firms still in basic AI application stages.
Personas in AI, friend or foe?
Are you using persona prompts with AI? Here’s what the research actually says. arxiv.org/html/2603… A new study from USC (“Expert Personas Improve LLM Alignment but Damage Accuracy”) tested expert persona prompts across six large language models and finally explains why the community has seen such mixed results. The finding is simple but important: persona prompts are an alignment tool, not a knowledge tool. When personas HELP: → Writing tone and style (scores jumped from 7/10 to 9/10 on professional email drafting) → Safety and refusal (jailbreak resistance improved by up to 17.
CodeWall says it hacked McKinsey’s AI platform. Here’s what holds up — and what doesn’t.
This reflects my personal assessment of publicly available reporting and CodeWall’s published blog post. I was not involved in the testing, I do not have access to McKinsey’s internal facts or forensic findings, and my views should be read as commentary and opinion rather than statements of verified fact.
A security startup called CodeWall claims its autonomous agent compromised McKinsey’s internal AI platform, Lilli, within two hours and gained unauthenticated read-write access to a production database containing tens of millions of consultant conversations. The vulnerability appears credible. The claimed scope of impact is not fully evidenced. The primary CodeWall post is here: codewall.ai/blog/how-… Independent reporting by Jessica Lyons in The Register is here: www.theregister.com/2026/03/0…
DeerFlow 2.0: ByteDance’s open-source AI agent harness for research and software tasks
DeerFlow 2.0, an open-source project from ByteDance, has quickly become one of the most visible AI agent releases of early 2026. The project’s public repository says it reached No. 1 on GitHub Trending on Feb. 28, 2026, and the repository currently shows about 25,000 stars and 3,000 forks. For teams evaluating agentic systems, DeerFlow deserves attention, but it also warrants disciplined review.
Heretic and the new reality of modifiable AI safety
Open-source large language models have made advanced generative AI broadly accessible. What is changing now is not only model capability, but the ease with which model behaviour can be altered after release — including behaviour that vendors and labs describe as “safety alignment.”
One of the most visible examples is Heretic, an open-source project that automates the removal of refusal behaviour in transformer-based language models. The project is not subtle about its purpose. It describes itself as “fully automatic censorship removal,” and it is gaining traction quickly.
This post does not provide instructions for disabling safeguards. Instead, it focuses on what is verifiably true about the tool, the research it is built on, and why this matters for security leaders, developers and governance teams.
Are Dorsey’s giant job cuts the start of an AI jobs apocalypse? Economists weigh in Block CEO Jack Dorsey’s decision to cut nearly half the company’s workforce raises questions about AI’s impact on jobs, but economists suggest this is a company-specific adjustment rather than a sign of a broader labor market shift. While AI may disrupt some jobs, experts like Claudia Sahm emphasize that it doesn’t necessarily lead to mass layoffs, and other economists believe AI will enhance productivity by changing workflows rather than eliminating jobs outright.
OpenAI is rolling out GPT-5.2 “Codex-Max” for some users OpenAI is rolling out GPT-5.2-Codex-Max, a new model for its Codex service, to select subscribers. This advanced version is expected to offer enhanced capabilities for long tasks, context management, and improved reliability, particularly with tool use and understanding visual inputs like screenshots.
The "Stein Standard": What the OpenAI ruling means for privacy and discovery
On Jan. 5, 2026, U.S. District Judge Sidney Stein affirmed a significant discovery order requiring OpenAI to produce 20 million de-identified ChatGPT conversation logs to plaintiffs in the consolidated copyright litigation involving The New York Times and other publishers.
As security and privacy professionals, we often warn about “Shadow AI” and data leakage. This ruling makes those risks concrete. Here is a balanced analysis of what happened and what it means for Canadian organizations.
French authorities investigate AI ‘undressing’ deepfakes on X French authorities are investigating AI-generated deepfakes on X after hundreds of women and teens reported non-consensual sexually explicit images created using the Grok chatbot. This investigation is part of an existing probe into X, with potential penalties including prison time and fines.
The "10 Per Cent" Myth: Why AI Capability Does Not Equal a Pink Slip
The headlines are everywhere, and they are designed to stop your scroll: “AI to Replace 1/10 of the Workforce.”
It is a terrifying number. It represents millions of livelihoods reduced to a statistic. But as a chief information security officer, I do not deal in headlines. I deal in risk, audits and rigorous data analysis.
When you strip away the hype and audit the primary sources released in late 2025—specifically from Project Iceberg (MIT), Yale and McKinsey—a completely different reality emerges.
We are confusing technical exposure with actual displacement.
Here is the fact-based reality of the AI labour market as we enter 2026.
The ‘Delete’ Button Is a Lie: A Canadian’s Guide to AI Data Retention
When you hit “delete” on a conversation with ChatGPT or Gemini, you likely expect it to vanish. In reality, that data often enters a digital limbo—accessible to the provider for 30 days, three years, or even seven years for certain safety-classifier metadata, depending on the fine print you didn’t read.
For paid subscribers, the assumption of privacy is dangerous. While corporate “Team” and “Enterprise” plans typically offer stronger contractual controls (including training restrictions and admin-managed retention), “Pro” and “Plus” users are frequently treated as consumers with slightly better perks, not better privacy.
China's open AI models are in a dead heat with the West
China’s open AI models are in a dead heat with the West - here’s what happens next www.zdnet.com/article/c… With the rising technological prowess and greater openness of Chinese models, the world is increasingly turning to the East for efficient and customizable AI, a new report finds. ZDNET’s key takeaways: Chinese AI models have caught up to US models in power and performance. China is leading in model openness. Much of the world may adopt the freely available Chinese technology.
Coursera to buy Udemy, creating $2.5 billion firm to target AI training | Reuters Coursera announced an all-stock deal to acquire Udemy, valuing the combined company at $2.5 billion. The merger aims to strengthen their position in corporate workforce training, particularly in AI, data science, and software development. The deal is expected to close in the second half of next year, pending regulatory and shareholder approvals.
Managing agentic AI risk: Lessons from the OWASP Top 10 | CSO Online The OWASP Top 10 for Agentic AI provides a framework to address the growing security risks associated with agentic AI adoption, offering practical guidance, threat taxonomies, and mitigation strategies for CISOs. While the list is immediately useful, some areas like detailed mitigation steps and attack likelihood require further development.
Microsoft Scales Back AI Goals Because Almost Nobody Is Using Copilot | Extremetech Microsoft has reportedly scaled back AI goals for its Copilot software due to low user adoption and sales, with some targets cut by 50%. While Microsoft disputes the sales quota claims, AI agents have shown low success rates in tasks, and Copilot lags behind competitors like ChatGPT and Google’s Gemini in market share.
I tested ChatGPT-5.2 vs Gemini 3.0 with 7 real-world prompts — here’s the winner | Tom’s Guide In a comparison of ChatGPT-5.2 and Gemini 3.0 across seven real-world prompts, ChatGPT-5.2 emerged as the overall winner, demonstrating superior emotional intelligence and psychological insight in its responses. While Gemini 3.0 excelled in specific areas like risk assessment and technical explanations, ChatGPT-5.2 consistently provided more human-like, wise, and grounding answers.
Autonomously Finding 7 FFmpeg Vulnerabilities With AI - ZeroPath Blog | ZeroPath This document details seven vulnerabilities found in FFmpeg, including buffer overflows and invalid frees, stemming from issues like integer truncation, unbounded serialization, off-by-one errors, and incorrect stream indexing. ZeroPath’s AI SAST identified these by analyzing allocation and copy alignment, framing invariants, packet builder capacities, cardinality propagation, and offset arithmetic integrity, often bypassing limitations of traditional fuzzers and static analysis tools.
Poetry can trick AI models like ChatGPT into revealing how to make nuclear weapons, study finds | The Independent A new study reveals that poetry-based prompts can trick AI models like ChatGPT into bypassing safety features and revealing instructions for creating malware or nuclear weapons. This method, termed adversarial poetry, successfully circumvented controls in major AI models, with poetic prompts leading to a significantly higher rate of unsafe replies compared to prose.
Australia Abandons Proposed Mandatory AI Rules in New Plan Australia has shifted from proposed mandatory AI rules to a voluntary framework, opting for existing laws on privacy and copyright instead of new AI-specific legislation. This decision has been met with support from business groups but criticism from academics and the Greens, who argue it lacks enforcement and adequate investment compared to international approaches.