Press Review — AI, Security and Token Economics — June 10, 2026
Enterprises are pushing AI-generated code to production despite documented security vulnerabilities, while Anthropic publicly acknowledges for the first time its silent behavioral restrictions in Claude Fable 5. Token economics take center stage: tokenmaxxing, the hidden cost of inference, and the formation of the Tokenomics Foundation all signal that AI governance is being built from the ground up.
Enterprises know AI-generated code is vulnerable; they’re shipping it anyway
A Checkmarx survey of 2,350 security leaders and developers across 14 countries is unambiguous: nearly half of production code is now AI-generated, and over a third of organizations leave half their known vulnerabilities unpatched for at least 90 days. The pressure for rapid ROI consistently overrides security discipline. Tools like Anthropic’s Mythos can surface vulnerabilities faster than human teams, but deployment velocity still outpaces remediation across the enterprise.
If Claude Fable stops helping you, you’ll never know
Simon Willison highlights a low-profile passage in the 319-page system card for Claude Fable 5 and Mythos 5: Anthropic now applies silent interventions to limit the model’s effectiveness for requests that could accelerate frontier AI development by competitors. These measures, via prompt modification and parameter-efficient fine-tuning, affect an estimated 0.03% of total traffic from under 0.1% of organizations. It marks the first time Anthropic has publicly documented covert enforcement aligned with its existing terms of service.
Another Linux Kernel Flaw Leads to Root Access Rights
CVE-2026-23111 is a use-after-free bug in the Linux kernel’s nf_tables code, rated 7.8 on the CVSS scale. It allows local unprivileged users to escalate to root via a syntax error, provided unprivileged user namespaces are enabled, which is the default on most distributions. The patch landed February 5th, but Exodus Intelligence and FuzzingLabs released technical details and proof-of-concept exploit code on June 8th. Systems that have not been restarted since before that date remain exposed.
Article in French.
Google Drops AI Plus Entry-Level Plan Price by Nearly 38%
Google has cut its AI Plus subscription from €7.99 to €4.99 per month, a 38% reduction, while doubling included storage from 200 GB to 400 GB across Drive, Photos, and Gmail. Usage caps in the Gemini app, Flow, and Antigravity are raised for subscribers, though limits now reset every five hours and scale with query complexity. AI Plus subscribers get twice the cap of non-paying users. Originally launched in September 2025, the plan is now available globally.
Article in French.
Siri AI Enters a New Era, Surpassing ChatGPT’s Shift to Gemini
Apple has pivoted Siri AI from OpenAI’s GPT-4o to Google’s Gemini models and Google Cloud infrastructure. The Private Cloud Compute architecture now extends to NVIDIA GPUs on Google Cloud, with Apple maintaining privacy compliance through specialized computing primitives. The system uses five hierarchical models: two run locally on devices, two in Apple datacenters, and one on Google Cloud for the most demanding queries. The shift reveals a deep strategic dependency on Google’s AI stack as Apple builds its own long-term capacity.
Article in French.
The tokenmaxxing backlash is coming
“Tokenmaxxing” describes the aggressive use of agentic AI tools like Claude Code to generate and fix code at scale. After an initial phase of stealth adoption, the practice became normalized across engineering teams within weeks. InfoWorld raises the pointed question: does the productivity gain actually justify the cost? Token spend climbs fast, code review workloads increase, and potential technical debt is hard to quantify. A useful read before committing to large-scale agentic development.
Beware of the genAI token trap
Tokens, the billing unit for every LLM interaction, are often treated as a technical footnote in AI projects. In practice, they represent the core cost driver of all generative AI services, and that cost scales directly with organizational adoption. InfoWorld warns that current pricing structures invite enterprises to underestimate future spend. Decision-makers who exclude token economics from long-term financial projections face significant budget surprises as AI usage matures across departments.
Meet Hades: The malware that lies to AI security agents
The Hades Campaign, uncovered by StepSecurity, is a supply chain attack targeting Python developer environments via popular packages like ensmallen. It deploys multi-layer payloads using the Bun toolkit for data exfiltration and lateral movement, combined with adversarial prompt injection to deceive AI-based security analysis tools. The broader signal: attackers now exploit AI security agents as an attack surface rather than just a defense layer. Security teams relying on AI-powered analysis tools should review their supply chain exposure.
Broadcom beefs up Spring security to protect against AI-enabled attacks
Broadcom is releasing the largest set of Spring security updates in the framework’s history, targeting AI-enabled attack vectors. The release includes a clean-room build architecture for Java dependencies to prevent supply chain contamination. Broadcom applies AI tools internally to identify vulnerabilities and validate patches. Tanzu Spring Enterprise customers receive zero-day access to validated CVE patches via the Spring Enterprise Repository before they reach the public open source channel.
10 MCP servers to connect LLMs with databases
Model Context Protocol (MCP) has established itself as the standard for connecting LLM-powered tools to databases and data sources. InfoWorld surveys ten MCP servers that enable natural language database operations, from lookups to record creation and administrative tasks, without writing SQL. Major platforms are covered, including Amazon Aurora (MySQL/PostgreSQL) via an official AWS MCP server. A practical starting point for engineering teams looking to give AI agents governed access to production data.
Linux Foundation Announces Intent to Launch the Tokenomics Foundation for Open AI Cost Standards
The Linux Foundation announces the formation of the Tokenomics Foundation, partnering with the FinOps Foundation to build open industry standards for AI infrastructure cost management. The scale of the challenge: global token usage is projected at 120 quadrillion tokens per month by 2030, with the inference market expanding to $255 billion. Eleven founding organizations, including Microsoft, Google Cloud, JPMorganChase, and Oracle, back the initiative. The goal is vendor-agnostic benchmarks for measuring token consumption and optimizing AI spending at enterprise scale.
