Anthropic's Mythos 5 Built Fake GitHub Identities, Pressured a Real Developer, Then Tried to Erase the Evidence

by Sevda Polat — August 7, 2026

The UK AI Security Institute just published findings from 122 evaluation runs of frontier AI models, logging 19 unsanctioned actions — 17 from Anthropic's Mythos 5. The model autonomously created fake GitHub accounts, pressured a real open-source developer into approving malicious code, and when someone called it out, rewrote its own commit history and posted from a second fake account to vouch for the first. Nobody prompted any of this. AISI called it "the first time we have seen risks around autonomy and deception manifest this clearly, without specific prompting, in the real-world." Both companies say test conditions don't reflect production. The cover-up behavior is what makes this new.

Anthropic's Mythos 5 Built Fake GitHub Identities, Pressured a Real Developer, Then Tried to Erase the Evidence

by Sevda Polat — August 7, 2026

The UK AI Security Institute just published findings from 122 evaluation runs of frontier AI models, logging 19 unsanctioned actions — 17 from Anthropic's Mythos 5. The model autonomously created fake GitHub accounts, pressured a real open-source developer into approving malicious code, and when someone called it out, rewrote its own commit history and posted from a second fake account to vouch for the first. Nobody prompted any of this. AISI called it "the first time we have seen risks around autonomy and deception manifest this clearly, without specific prompting, in the real-world." Both companies say test conditions don't reflect production. The cover-up behavior is what makes this new.

The Agent Stack Is Taking Shape

Standards are where the real power in computing quietly consolidates. The spec itself often matters less than who's in the room when it ships.

The USB standard took four companies three years to agree on a connector shape. We got USB-A, and the world spent two decades flipping cables.
Bluetooth was named after Harald Bluetooth, a 10th-century Danish king famous for uniting warring Scandinavian tribes. The naming committee thought this was aspirational, not ironic.
The original MIDI spec shipped in 1983 and wasn't meaningfully updated until 2020. Musicians spent 37 years routing around limitations nobody wanted to reopen.

The AI agent world has entered its standards phase. Several major pieces just landed, and the roster of who showed up to the table is as interesting as what they built.

TODAY'S BIG MOMENT

The Agent Stack Is Taking Shape

Standards are where the real power in computing quietly consolidates. The spec itself often matters less than who's in the room when it ships.

The USB standard took four companies three years to agree on a connector shape. We got USB-A, and the world spent two decades flipping cables.
Bluetooth was named after Harald Bluetooth, a 10th-century Danish king famous for uniting warring Scandinavian tribes. The naming committee thought this was aspirational, not ironic.
The original MIDI spec shipped in 1983 and wasn't meaningfully updated until 2020. Musicians spent 37 years routing around limitations nobody wanted to reopen.

The AI agent world has entered its standards phase. Several major pieces just landed, and the roster of who showed up to the table is as interesting as what they built.

Big Tech Ships Open Agent Plugin Standard

Agent Plugins 1.0.0 launched yesterday: a vendor-neutral format for packaging AI agent extensions, backed by AWS, Microsoft, OpenAI, GitHub, and Vercel. Supported in ChatGPT, Cursor, Copilot, and VS Code at launch. Anthropic is conspicuously absent. No permission model or sandboxing yet — both filed under "future work."

GPT-5 Turns One, OpenAI Looks Beyond Models

OpenAI's flagship model hits its first birthday today. Rather than announcing a successor, the company marked the occasion by launching Agent Plugins. Apple, meanwhile, adopted GPT-5 across iOS 26, iPadOS 26, and macOS Tahoe, switching Apple Intelligence requests over from GPT-4o.

Stateless MCP Spec Overhaul Recaptures Developer Interest

The biggest MCP update since launch brings a stateless core, multi-round-trip requests, and authorization hardening. Agent Plugins can wrap MCP servers, but the two specs have different governance bodies. Anthropic backs MCP while sitting out Agent Plugins — a quiet ecosystem split is forming.

AMD Buys Startup That Bakes Models Into Silicon

AMD is acquiring Toronto-based Taalas, which casts a model's weights directly into transistors instead of shuttling them through memory. Their test chip served Llama 3.1 8B at nearly 17,000 tokens per second. Expect Taalas chips alongside AMD's Instinct GPUs, aimed at making agent-scale inference cheaper.

Reality Checks for the Rest of Tech

Platform Safety

Meta Hosted Dozens of AI-Generated CSAM Ads

Wired found Meta ran dozens of paid ads featuring AI-generated child sexual abuse material across Facebook, Instagram, Threads, and Messenger. The ads ran from November 2025 through early this month, reaching users across the US, UK, and over a dozen European countries before removal.