Perplexity Announces “Search as Code” — A New Approach Where AI Writes Its Own Search Pipelines
Perplexity has announced a new architecture called “Search as Code.” The approach does away with calling search APIs as fixed endpoints, instead having AI models write their own search routines in Python.
Under this system, agents autonomously handle filtering and deduplication within a sandbox. As a result, it reportedly achieved scores surpassing OpenAI's and Anthropic's systems on major benchmarks. It's also worth noting that token costs were reduced by up to 85%.
In traditional RAG (Retrieval-Augmented Generation) pipelines, each step — generating search queries, filtering results, and integrating them into context — was fixed. “Search as Code” flips this premise, delegating the design of the search process itself to the agent. It's fair to say this is an attempt to push AI agent autonomy to a whole new level.
That said, validation of the safety and reproducibility of agent-written code in production appears to still be a work in progress. While the cost savings are attractive, we'd like to wait for community verification of execution overhead within sandboxes and behavior in edge cases.
ChatGPT Gains “Lockdown Mode” — A New Feature to Counter Prompt Injection
OpenAI has added a new security feature called “Lockdown Mode” to ChatGPT. When enabled, this mode disables web access, Deep Research, and Agent Mode entirely.
The goal is to prevent the theft of sensitive data through prompt injection attacks. Prompt injection is an attack technique that rewrites an AI's instructions via malicious text, triggering unintended behavior — and it remains a problem for LLMs with no fundamental solution in sight.
Lockdown Mode works by blocking the final stage of the attack chain, namely data exfiltration. However, OpenAI itself explains that the mode does not completely prevent attacks and serves merely as a last line of defense. It's important to note that it does not fundamentally prevent prompt injection itself.
For enterprise ChatGPT use, this looks like an effective mitigation when handling sensitive information. However, since it sacrifices web access and agent features, users will need to judge the trade-off between convenience and security for themselves.
LLM Crawlers Cripple SourceHut — AI's Shadow Weighs on Infrastructure
SourceHut, a software development platform, suffered a service outage due to a surge of traffic from crawlers collecting training data for LLMs (large language models).
SourceHut is known as a lightweight open-source development platform, but a rapid increase in crawlers harvesting code and documentation as LLM training data reportedly pushed its server load past the limit. This shows that LLM crawlers can pose a serious threat to smaller services with limited resources, unlike large platforms such as GitHub.
Similar problems have been reported on other small-scale web services, and the impact of AI companies' data collection on infrastructure is becoming an industry-wide challenge. robots.txt-based controls are not always respected, and the development of technical and legal countermeasures is considered urgent.
UK Plans Strategic Procurement from Domestic AI Chip Makers
The UK government has announced plans to strategically procure AI chips from domestic semiconductor companies. Technology Secretary Liz Kendall is reportedly set to unveil a policy of making “strategic purchases” from UK-based companies in a speech at London Tech Week.
The move appears to be driven by concerns that the supply of AI semiconductors is concentrated among a handful of companies, including NVIDIA. The UK is home to its own chip ventures such as Graphcore, and the government appears determined to build a homegrown supply chain from the standpoints of nurturing domestic industry and national security.
However, details remain unclear about the competitiveness of UK-made chips and the actual scale of procurement — whether this stays a mere declaration or becomes an effective policy remains to be seen.
Hugging Face Releases “Her,” a Tool for Analyzing Claude Code Sessions
At a Hugging Face hackathon, a tool called “Her (हेर)” for analyzing Claude Code sessions was announced. The name means “to look” in Hindi.
In development work with Claude Code, there have been limited ways to look back over an entire session and see what decisions the agent made when generating code. Her addresses this challenge by providing session analysis and visualization. Details of its specific features are expected to be revealed soon, but it's an intriguing effort from the perspective of improving transparency in AI coding.
---