|
|
|
|
|
GitHub Agentic Workflows grants the AI agent cross-repo read access, and it does not wall off system instructions from a user-supplied issue body, so a crafted issue acts as indirect prompt injection.
|
|
JULY 10TH, 2026 · BY JONATHAN HILDEBRANDT
|
|
Noma Security researcher Sasi Levi published GitLost on July 7: anyone who can open an issue in a public repo can pull private repo contents from the same GitHub org, with no credentials and no code. GitHub Agentic Workflows grants the AI agent cross-repo read access, and it does not wall off system instructions from a user-supplied issue body, so a crafted issue acts as indirect prompt injection. Levi's fake VP-of-Sales issue got the agent to fetch private README files and post them publicly. GitHub had no patch as of July 10.
Why it matters. Any org with Agentic Workflows enabled is exposed right now, against default config. Restrict agent scopes to read-only on private repos, or disable issue-triggered agents until a fix ships. Our updated Copilot review has the admin permission checklist: run the audit. Attack writeup: Noma Security.
|
|
Run the admin permission audit →
|
|
|
GPT-Live-1 does full-duplex voice
JULY 10TH · PONDERO NEWSDESK
OpenAI shipped GPT-Live-1 and GPT-Live-1 mini on July 8. Where every prior voice system stops listening while it talks, GPT-Live-1 processes incoming and outgoing audio at the same time and routes hard queries to GPT-5.5 in the background. If you build voice agents, interruption handling stops being your problem to hack around. OpenAI.
|
GPT-5.6 Sol, Terra, and Luna hit Copilot
JULY 10TH · PONDERO NEWSDESK
GitHub added all three tiers on July 9. Terra and Luna run on any paid plan including base Pro; Sol, the long-agentic-task tier, is gated to Pro+, Max, Business, and Enterprise, and Business or Enterprise admins must enable it. If you want the reasoning tier, check whether your admin has flipped it on. GitHub changelog.
|
SK Hynix opens on Nasdaq at $158.14
JULY 10TH · PONDERO NEWSDESK
SK Hynix (SKHYV) began when-issued Nasdaq trading on July 10 at $158.14 per ADS, closing a $28.1 billion offering that tops Alibaba's $21.9 billion 2014 record for the largest foreign-company US equity debut, with $7 billion committed by three cornerstone investors before it opened. All proceeds go to HBM production, and SK Hynix supplies the HBM inside Nvidia H100 and H200 GPUs, so this is the memory bottleneck raising cash. Price: stockanalysis.com. Offering size: Proactive.
|
Kimi K2.7 is the first open-weight model in Copilot
JULY 10TH · PONDERO NEWSDESK
GitHub finished the Kimi K2.7 Code rollout to Business and Enterprise on July 7 at $0.95 input and $4.00 output per 1M tokens under Usage Based Billing, well under Claude Fable 5 rates. It is off by default and admins must opt in. The catch: Moonshot AI is a Chinese company, so your compliance team gets a vote before you enable it. We cover it in the updated Copilot review. Pricing: GitHub changelog.
|
Claude Reflect scores how you use AI
JULY 10TH · PONDERO NEWSDESK
Anthropic launched Claude Reflect on July 9 for Free, Pro, and Max: a breakdown of your topics, task types, and timing, scored against a 4D Fluency Framework (Delegation, Description, Discernment, Diligence), with quiet-hours and break nudges alongside. TechCrunch's read is that the dashboard doubles as a retention hook that frames more usage as progress, and Anthropic has not said what syncs to its servers. Worth a look, worth a skeptical one. Anthropic, TechCrunch.
|
Cloudways
JULY 10TH · PONDERO NEWSDESK
Managed cloud hosting for the self-hosted AI stack: n8n, Open WebUI, and MCP servers without babysitting a raw VPS. Live in a few minutes. Try Cloudways →
|
Make
JULY 10TH · PONDERO NEWSDESK
Visual, debuggable workflow automation for AI agents, with a clear run history when a scenario breaks. Try Make →
|
|
•
|
Fable 5 free access extended through July 12. Anthropic reversed the July 8 billing cutoff; Pro, Max, Team, and Enterprise get five more days. After July 12, usage credits kick in at $10 and $50 per million tokens (per Anthropic's pricing page).
|
Copilot re-rated to 4.4
JULY 8TH · PONDERO REVIEW
Six GA launches in one week - from vision to Kimi K2.7 to cost centers - moved our Copilot score to 4.4 with fresh per-plan picks. Read it →
|
Make AI Agents vs Lindy
JULY 6TH · PONDERO COMPARISON
Make gives you a visual, debuggable workflow; Lindy acts like a teammate on your inbox and calendar. Which one fits your ops stack. Read it →
|
Claude Sonnet 5 vs GPT-5.5 vs Gemini 3.1 Pro
JULY 5TH · PONDERO GUIDE
Sonnet 5 is cheapest at $2 per million input through August 31 (per Anthropic's pricing) and leads SWE-bench Pro at 63.2 percent (per SWE-bench benchmark data), with per-persona picks and cited pricing. Read it →
|
|
|
|
|
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.
X
·
Bluesky
·
LinkedIn
|
|
