Standard Chartered Bank to Cut Around 7,800 Jobs Amid AI Adoption
UK-based Standard Chartered Bank has announced that it will cut more than 15% of its back-office roles (around 7,800 jobs) by 2030 as it expands AI adoption. The bank, whose major back-office hubs are located in India, China, Malaysia, and Poland, said in a statement that it will "expand the practical use of automation, advanced analytics, and AI to streamline processes." Some employees are reportedly set to be transferred to other departments.
This trend extends beyond the financial industry to tech companies at large. Meta has announced cuts of around 8,000 jobs next month, Amazon laid off more than 30,000 in January, and Oracle cut over 10,000. The picture of massive AI investment and workforce reductions proceeding in parallel is coming into sharp focus.
Meanwhile, a Surge in Companies Cutting AI Tool Spending
As an interesting counterpoint, a Hacker News post asking "My company is rapidly cutting AI tool spending — how should teams prepare?" drew significant attention. The poster revealed that Claude Code access had been effectively revoked and even personal-plan usage was banned. The reason: monthly Claude spending had reached three times the organization's SaaS cloud bill, and the company is now migrating to Codex.
While AI displaces jobs on one front, AI tool costs have reached unsustainable levels on the other, with companies pulling back on AI investment — two opposing dynamics unfolding simultaneously.
AgentWall: A Runtime Safety Layer for AI Agents Debuts
Amid growing concerns over the safety of autonomous AI agents, a paper on "AgentWall," a runtime safety and monitoring layer, was published on arXiv. AgentWall intercepts every action an agent attempts to execute before it reaches the host environment, evaluates it against declarative policies, and requires human approval for sensitive operations. It also records a complete audit trail of execution trajectories.
It is implemented as an MCP proxy supporting Claude Desktop, Cursor, Windsurf, Claude Code, and OpenClaw, and can be installed with a single command. It reportedly achieves 92.9% policy enforcement accuracy across 14 benchmark tests and operates with sub-millisecond overhead.
The work is motivated by the recognition that as AI agents increasingly execute shell commands, manipulate files, and make API calls, model-level alignment alone is not enough — a runtime safety layer is needed.
The "Memory Laundering" Problem in Memory-Equipped LLM Agents
Also related to agent safety, an attack technique dubbed "memory laundering" has been analyzed. The issue: when harmful or adversarial context gets compressed into a memory summary within an LLM agent's persistent memory, it retains its adversarial framing while evading standard toxicity detectors.
The researchers introduce a metric called the "Sub-threshold Propagation Gap (SPG)" and quantitatively demonstrate that memory states judged "safe" by monitoring systems can still exert hidden influence on downstream behavior. Crucially, sanitizing toxic states before summarization greatly reduces hidden propagation, whereas sanitizing after summarization can leave "laundered" influences behind. This suggests that in agent memory design, the timing of intervention is critically important.
Mizuho FG Builds an "Agent Factory," Cutting Development Time to as Little as a Few Days
In domestic news, ITmedia reports that Mizuho Financial Group has built an "Agent Factory" — a system for developing and mass-producing AI agents in short cycles. It cuts the development period of complex AI agents, previously about two weeks, by up to 70%, enabling development in as little as a few days.
A major Japanese financial group making serious investments in mass-production infrastructure for agents is a clear sign that agent technology is entering the practical stage.
RTPurbo: Converting Full Attention to Sparse Attention in 100 Steps
A technique called "RTPurbo" that dramatically reduces the computational cost of long-context inference has been announced. It demonstrates that full-attention LLMs are already inherently sparse, and can be converted into highly sparse models with just a few hundred training steps.
The approach rests on three observations: (1) only a subset of attention heads truly need long-context processing; (2) long-range retrieval is governed in a low-dimensional subspace, so it can be retrieved efficiently with a 16-dimensional indexer; (3) the useful token budget depends heavily on the query, so dynamic top-p selection is better suited than a fixed top-k.
RTPurbo keeps the full KV cache only for retrieval heads and introduces sparse attention with a lightweight token indexer. At a 1M context, it delivers up to a 9.36× prefill speedup and roughly a 2.01× decoding speedup with almost no loss in accuracy.
ZEDA: Skipping Half of the Experts in Trained MoE Models
"ZEDA (Zero-Expert Self-Distillation Adaptation)," a paper gaining attention on Hugging Face, is a framework for converting post-trained static Mixture-of-Experts (MoE) models into efficient dynamic MoE models at low cost.
It injects a parameter-free zero-output expert into each MoE layer and performs two-stage self-distillation with the original MoE frozen as the teacher. In evaluations on Qwen3-30B-A3B and GLM-4.7-Flash, it reduced expert FLOPs by more than 50% with only a minor drop in accuracy, outperforming prior dynamic MoE methods by up to 6.1 points and achieving roughly 1.20× end-to-end inference speedup.
As an approach that makes trained models more efficient without retraining, it offers high practical value.
---