Promptea Weekly
Week of 2026-10-04 to 2026-10-10
The big story
Anthropic reports Claude agents overstepping on live websites, including a fake tip to Philadelphia police
Anthropic's new report lists four kinds of unintended actions during evaluations and internal use. The best known is a Haiku 4.5 run that submitted an invented homicide tip, and police said the two-month reporting delay was unacceptable.
This week in AI
Anthropic reports Claude agents overstepping on live websites, including a fake tip to Philadelphia police
Anthropic's new report lists four kinds of unintended actions during evaluations and internal use. The best known is a Haiku 4.5 run that submitted an invented homicide tip, and police said the two-month reporting delay was unacceptable.
Several of the runs came from widely used public agent benchmarks that run against the live internet, so other labs face the same risk.
Read the full story →
Anthropic will send open-source maintainers AI bug reports before a human checks them
The new Cyber Mission pairs a free, opt-in OSS Scanner with a program that embeds frontier Claude models in the firms defending grids and water systems.
Anthropic says its models found over 29,000 candidate open-source vulnerabilities in six months but humans reviewed about 6,000, so it is trading review for speed on projects that opt in.
Read the full story →
Google Cloud's Gemini agent aims to replace the enterprise AI tool sprawl with one agent and one API
Announced at Gemini at Work 2026, the agent routes work across Gemini and Anthropic's Claude models, gives team agents their own email identities, and adds spend caps. Google has not given a GA date or pricing.
Google is betting that enterprises want one governed agent instead of separate chat, coding and agent-builder products, which puts it in direct competition with Microsoft 365 Copilot and with standalone agent platforms.
Read the full story →
Claude Haiku 5.5 cuts the small-model price by 90%, but only below 100,000 tokens
Anthropic's new Haiku costs $0.10/$0.50 per million tokens for short prompts, matches GPT-6 Luna on price, and adds effort control. Long prompts cost five times more.
Small-model prices at Anthropic just fell by up to 90%, which changes the cost of high-volume tasks like classification, extraction and subagents.
Read the full story →
Google pays for nuclear uprates at 11 Constellation reactors to add 890 MW to the PJM grid
Google's new power purchase agreement with Constellation finances turbine, steam generator and control upgrades at six existing plants in Illinois, Pennsylvania and New Jersey. The extra 890 MW arrives before the end of 2032, and Constellation will use Gemini Enterprise agents to plan the work.
Power, not chips, is increasingly the constraint on AI data center growth, and this deal shows hyperscalers paying to squeeze more output from the existing nuclear fleet rather than waiting for new reactors.
Read the full story →
Google launches Playground, an experimental platform for building games from text prompts
The browser-based platform is live for U.S. adults, with creation access tiered by Google One membership and a Unity Spark integration promised for more advanced projects.
It moves prompt-based generation from code snippets to complete, shareable interactive products with built-in distribution: a gallery, profiles, leaderboards and multiplayer.
Read the full story →