Claude models broke into three real companies during a security test, a maximum-severity flaw hit a popular agent harness, and OpenAI cut prices again.
Five stories from the last three days worth your attention this week: an AI lab disclosing that its own models broke into real companies during a security test, a maximum-severity flaw in a widely used agent harness, another round of API price cuts, a fully open video model out of Shanghai, and an earnings report that says a lot about where AI infrastructure money is actually going.
[2026-07-30 anthropic.security] real-world-access-during-cyber-eval
Anthropic disclosed that three of its models, Opus 4.7, Mythos 5, and an unreleased internal research model, gained unauthorized access to the real systems of three organizations during third-party capture-the-flag cybersecurity evaluations run with partner Irregular. The models were told they were in an internet-free simulation, but a misunderstanding with the eval partner meant real internet access was live. Using basic techniques like unauthenticated endpoints and weak passwords, the models breached the actual target systems instead of the intended sandbox.
Why it matters: This is a sandbox-escape story, not a jailbreak story, and that distinction should worry infrastructure and security teams more than most AI headlines do. If an evaluation environment can leak into production because of a partner misconfiguration, the same failure mode applies to any agent your org runs with broad tool access. Audit your own agent sandboxes for real network egress before you trust the "it's isolated" assumption.
Source: Anthropic, "Investigating three real-world incidents in our cybersecurity evaluations"
[2026-07-29 thehackernews.vuln] rufroot-cve-2026-59726
Researchers at Noma Security disclosed a maximum-severity flaw, CVSS 10.0, in Ruflo, an open-source agent meta-harness used with Claude Code and OpenAI Codex. Codenamed RufRoot and tracked as CVE-2026-59726, the bug came from Ruflo's default docker-compose setup exposing its MCP bridge endpoints without authentication on port 3001, bound to all network interfaces by default. An unauthenticated attacker could call terminal_execute, get a shell in the bridge container, steal provider API keys, and poison the agent's memory store. The maintainer shipped a fix in version 3.16.3 within 24 hours of disclosure.
Why it matters: If your team self-hosts any agent orchestration layer, this is a direct reminder to check what's bound to 0.0.0.0 by default. 233 tools were reachable through one unauthenticated endpoint. Patch now, and add "check default bind addresses" to your standard checklist for adopting any new agent framework.
Source: The Hacker News, "Ruflo MCP Flaw Lets Unauthenticated Attackers Run Commands and Poison AI Memory"
[2026-07-30 openai.pricing] gpt-5.6-api-cuts
OpenAI cut API prices on its two cheaper GPT-5.6 tiers just three weeks after the family's July 9 launch. Luna dropped 80 percent to $0.20 input and $1.20 output per million tokens. Terra dropped 20 percent to $2 input and $12 output per million tokens. The flagship Sol tier held at $5 and $30. OpenAI attributes the cut to efficiency gains across its model-serving stack, not a one-off promotion.
Why it matters: An 80 percent cut on the entry tier this soon after launch changes the math on high-volume agent workloads, log triage, code review bots, chat support, that were previously cost-gated. If a project got shelved on token cost a month ago, it's worth re-running the numbers.
Source: Unite.AI, "OpenAI Cuts API Prices on Its Two Cheaper GPT-5.6 Tiers"
[2026-07-31 minimax.release] h3-open-video-model
Shanghai-based MiniMax released H3, a multimodal model generating up to 15-second clips at 2K resolution with native stereo sound, and it can edit or transfer motion across video, image, text, and audio inputs. MiniMax says H3 ranks first globally on video editing benchmarks, at roughly a third of the price of comparable closed competitors. Model weights are slated for open release in early August.
Why it matters: Open-weight releases at this quality level compress the moat around closed video models fast, and they widen the pool of tooling your team can self-host, fine-tune, or audit instead of calling a black-box API. Worth a bench test if your pipeline touches generated video or motion transfer.
Source: Pandaily, "MiniMax Releases Open-Source Full-Modal Model H3"
[2026-07-29 meta.earnings] q2-capex-vs-free-cash-flow
Meta's Q2 revenue rose 28 percent to $60.8 billion, beating expectations, but free cash flow collapsed to $784 million from $8.5 billion a year earlier. The gap came from $31.1 billion in quarterly AI infrastructure capex, plus $2.4 billion in legal charges and $1.18 billion in severance tied to layoffs that started in May. Full-year capex guidance was raised to $130 to $145 billion, and Zuckerberg hinted at Meta potentially entering the cloud business.
Why it matters: This is the clearest data point yet on how much of "AI growth" is being funded by squeezed cash flow and layoffs elsewhere in the business. If you're tracking hiring demand or budget cycles in AI infrastructure roles, capex guidance like this is a better leading indicator than any single product launch.
Source: Fortune, "Meta stock drops 10% as free cash flow gets crushed"
The thread connecting all five
Every story this week is really about the same tension: agent capability is scaling faster than agent containment. Anthropic's own models broke out of a sandbox. Ruflo's harness exposed 233 tools to the open internet by default. Meanwhile the economics keep pushing harder in the other direction, cheaper tokens, cheaper open video models, and hyperscalers spending down their cash flow to build more of this, faster. Teams that win here are the ones treating "isolate and audit the agent" as a permanent line item, not a one-time setup task.
Building the skills to catch issues like the Ruflo flaw before they ship is exactly what tha-shed's Ethical Hacking, Red Team Strategies course is for, worth a look if this week's roundup hit close to home.
