AI News Briefing — June 8, 2026

A new study using models from 4 million to 4 billion parameters explains why small models fail at rare tasks: frequent tasks overwrite learned skills. The researchers suggest increasing how often a target task appears in training data may be as effective as scaling up model size.

Perplexity introduced Search as Code, an architecture that lets AI models write their own search routines in Python within a sandbox instead of calling fixed APIs. The approach reportedly beats OpenAI and Anthropic on key benchmarks while cutting token costs by up to 85 percent.

Researchers from UIUC, UC Berkeley, and Chroma released Harness-1, a 20B retrieval subagent built on gpt-oss-20b and trained with reinforcement learning inside a stateful search harness that handles bookkeeping while the policy makes semantic decisions. The weights and harness code are publicly released.

OpenAI reportedly plans its biggest ChatGPT overhaul since launch, rebuilding it as a superapp bundling coding tools, AI agents, and partner apps like Canva and Booking.com. Internally the company says chat is dead and is betting on autonomous task-handling agents.

OpenAI launched a Lockdown Mode for ChatGPT that disables web access, Deep Research, and Agent Mode to reduce data theft via prompt injection attacks. The feature only blocks the final exfiltration step and does not fully solve prompt injection, which remains unresolved.

The Guardian uses six charts to map the AI boom, noting SpaceX seeking a $1.77tn valuation, Anthropic filing for an IPO, and OpenAI expected to follow amid massive datacenter spending. The piece weighs accelerating consumer adoption against questions about whether returns will justify the multitrillion-dollar investment.

Anthropic hired Clive Chan, by his account OpenAI's second hardware employee in its custom chip program, who also worked on Tesla Autopilot ASICs and the OpenAI-Broadcom partnership. The move comes as both firms prepare IPOs and Anthropic reportedly weighs developing its own AI chips.

OpenAI is reportedly continuing work on a super app, with a senior employee declaring chat is dead. The framing signals a strategic pivot toward an agent-centric product beyond the chatbot interface.

DeepSeek topped Ramp's trending software vendors in June 2026 as US companies adopt cheaper AI and send data directly to the paid service. Ramp's economist cites cost awareness as the driver while warning of security risks of using Chinese models.

A survivor of a January 2025 Nashville school shooting is suing Omnilert, maker of an AI gun detection system that failed to spot the handgun used in the attack. The lawsuit alleges the company knew or should have known about operational limitations like camera angle, lighting, and weapon visibility that could cause detection failures.

The article examines a rising wave of anti-AI and anti-tech extremism, including an arrest of a man who allegedly plotted to burn OpenAI headquarters and Sam Altman's house, plus other ecofascist and Unabomber-inspired plots. It frames AI backlash as an emerging driver of political violence.

An investigation describes how AI shopping assistants like ChatGPT recommend fake websites impersonating legitimate retailers, leading buyers to scam sites. The phenomenon points to a poisoning vector where bad actors manipulate AI tool recommendations.