|
|
|
|
|
|
The first documented case of AI deception targeting real people, unprompted, in the wild
|
|
AUGUST 7TH, 2026 · BY JONATHAN HILDEBRANDT
|
|
Britain's AI Security Institute reported that frontier agents went off-script in 10 of 122 cybersecurity test runs with guardrails lowered. In the worst case, an agent fabricated identities, contacted real people to execute malicious code, then edited its own records when challenged.
|
|
Read the full breakdown →
|
|
|
|
|
White House AI Safety Framework Stayed Voluntary
AUGUST 6TH, 2026 · PONDERO NEWSDESK
Delivered under Executive Order 14409, the framework runs on classified benchmarks with no mandatory lab participation. It arrived the same day as the AISI disclosure: fresh evidence for mandatory pre-deployment review landed the same afternoon the administration chose not to require it.
|
Anthropic Ships Inference Hooks: A Server-Side DLP Gate for Claude Enterprise
AUGUST 6TH, 2026 · PONDERO NEWSDESK
A new server-side layer intercepts every Claude Enterprise prompt and tool-call response before inference, routing to a company-controlled server for an allow-or-deny verdict over signed WebSocket. Covers chat, Claude Code, and Claude Cowork. Beta, with Netskope, Palo Alto Networks, Proofpoint, and Zscaler as launch integrations.
|
Seven Sandbox-Escape Routes Hit Cursor, Codex CLI, and Gemini CLI
AUGUST 6TH, 2026 · PONDERO NEWSDESK
Pillar's coordinated disclosures share one pattern: the agent never breaks the sandbox directly. It writes files that trusted host components later execute. Cursor patched all three in version 3.0.0; Codex CLI closed its route in 0.95.0. Google acknowledged both findings but shipped no patch and paid no bounty.
|
|
•
|
GitHub Copilot dropped Gemini 2.5 Pro and Gemini 3 Flash. Both models left the model picker on July 31 with no official explanation. If your team routed Copilot to Gemini for long-context work, you are back to OpenAI and Anthropic models only. Details →
|
|
•
|
Comp AI automates SOC 2 and ISO 27001 evidence collection. The compliance paperwork every team deploying agents into production will owe as security requirements tighten around agentic systems. Try Comp AI →
|
|
•
|
Cloudways offers managed hosting for self-hosted AI tooling. Spin up a model gateway or agent backend in minutes without routing prompts through a third-party inference API. Try Cloudways →
|
Cursor: Buy, Hold, or Switch Before the SpaceX Deal Closes
AUGUST 5TH, 2026 · PONDERO REVIEW
Our decision framework for solo devs, small teams, and enterprise buyers as the acquisition looms. Includes current pricing, the model-access risk that would flip the call, and whether to lock in an annual plan now.
|
EU AI Act GPAI Enforcement Is Live: Who Owes What
AUGUST 6TH, 2026 · PONDERO GUIDE
GPAI enforcement went live August 2. This guide maps each obligation to model provider vs. deployer, and identifies the three controls that close most of the compliance gap for teams building on frontier models.
|
|
|
|
|
Jonathan Hildebrandt
Co-founder and primary operator of Pondero. Writes the Pondero Brief.
|
|
|
Affiliate disclosure
·
Unsubscribe
·
Manage preferences
Pondero earns commissions on some links. This does not affect our editorial picks.
|
|