Need to know
- GitHub Copilot ships Agent Plugins 1.0 GA, betting on model portability over any single model.
- SentinelOne opens Purple AI autonomous threat investigation to all Singularity customers at no extra charge.
- Palo Alto Networks is reportedly in takeover talks for CyberArk, sending CyberArk shares up 18%.
- DeepSeek open-sources its agent harness under MIT license, earning 33,000 GitHub stars within hours.
- Gartner warns agentic workflow inference costs will increase more than fivefold through 2028.
New Releases
GitHub released Copilot Agent Plugins 1.0 into general availability today, pairing it with a redesigned CLI and multi-model switching built into the core product.
- Model portability is the explicit strategy: the update is architected so developers can swap underlying models mid-workflow rather than commit to one, a direct response to fragmentation across GPT-5.6, Claude, and Gemini variants.
- CLI redesign ships simultaneously, adding a tabbed interface and zero-config setup that reduces the friction cost of adopting the agent-mode features.
SentinelOne made its Purple AI Agentic Investigation capability available to all Singularity Platform customers today, offered as a complimentary trial with zero-click autonomous threat response.
- Zero-click investigation means Purple AI detects, investigates, verifies, and initiates response when a threat crosses a defined threshold, without waiting for analyst input — directly addressing SOC analyst capacity constraints.
- Timing is deliberate: the move comes as CrowdStrike and SentinelOne both hit 52-week stock highs last week on platform adoption momentum, and ahead of SentinelOne's August 28 earnings.
DeepSeek released DeepSeek Harness v0.1 in developer preview under the MIT license today, shipping an agent runtime where every component — model adapter, tool registry, session log, and control loop — is a replaceable plugin.
- Everything-is-a-plugin architecture means the harness is model-agnostic by design; enterprises can swap in any model without rewriting the orchestration layer, which is a direct threat to proprietary harness lock-in.
- 33,000 GitHub stars within hours of release signals the developer community is hungry for a vendor-neutral, open runtime — the same dynamic that made LiteLLM's proxy layer a de facto standard.
Gartner published research today predicting that AI inference costs per agentic workflow will increase more than fivefold through 2028, warning that falling token prices are masking rising total AI spend.
- The core argument: each successive generation of agentic products adds workflow steps and model calls faster than token efficiency improves, so cheaper tokens paradoxically fund more expensive total runs.
- Implication for product leaders: inference cost management — not feature parity — is now a primary margin variable, and organizations building agentic products on today's pricing assumptions will face a structural cost problem within 24 months.
Funding
CyberArk shares jumped as much as 18% today on reports that Palo Alto Networks is in acquisition talks, a deal that would fold the market's leading privileged access management platform into the broadest enterprise security stack — and almost certainly trigger FTC scrutiny given Palo Alto's dominant position across endpoint, network, and cloud security.
Nvidia launched the Open Secure AI Alliance today with close to 40 members including Microsoft and IBM but notably excluding OpenAI, Anthropic, Google, and Meta — a coalition framing AI security tooling as open infrastructure, which benefits Nvidia's hardware-neutral positioning while sidelining the closed-model labs.
Reporting circulating today claims Anthropic showed investors a path to $190-200B in annual revenue by 2028 and is targeting a September-to-early-October IPO window at a valuation near $965 billion — which would make it the largest software IPO in history and reset comparables for every AI-native company behind it.
Case Studies
Jamf Threat Labs today published research on AmnesiaStealer, a new macOS information-stealing malware that hijacks active browser sessions via remote control and spreads through ClickFix social engineering attacks using a counterfeit GitHub page with a forged Verified Publisher badge.
- The delivery mechanism is novel: the fake GitHub page is visually identical to the real one including the Verified Publisher badge, which bypasses user skepticism trained on obvious phishing indicators.
- Enterprise exposure is real: browser session hijacking without credential theft means the attacker inherits authenticated SaaS sessions — Okta, Salesforce, GitHub — without triggering password-based detections.
Trending on X
- Persistent vs. ephemeral AI compute Ethan Mollick and others are mapping the emerging split in how labs give AI agents a computer: Claude Code and Codex use your local machine, ChatGPT Work spins up a one-time cloud machine, and Grok Bot gives each agent a persistent cloud PC — and the community is debating which model wins for long-running enterprise workflows.
- DeepSeek harness open-source reception Developers are reacting to DeepSeek's MIT-licensed agent harness with significant enthusiasm, drawing comparisons to LiteLLM's rise and debating whether a composable, model-agnostic runtime from a Chinese lab creates supply-chain risk for enterprise adopters.
- GLM-5.3 cybersecurity benchmark claims Zhipu's claim that GLM-5.3 beats Claude Fable 5 and GPT-5.6 Sol on the CyberGym vulnerability discovery benchmark is circulating among security engineers, with debate over whether CyberGym is a meaningful real-world proxy or a tailored benchmark that flatters Chinese models.
- Agentic AI cost accounting gap Following the Gartner fivefold cost warning, practitioners on X are sharing that most enterprise agentic pilots are tracked by capability demos rather than cost-per-task, with several noting that no one in their org has modeled what the workflow costs at production volume.
- AI agent harness consolidation race The same week Microsoft's Agent Harness hit GA and DeepSeek open-sourced its harness, developers are debating whether the orchestration layer is consolidating around a small number of frameworks or fragmenting further — with the open-source camp citing DeepSeek's star count as evidence the market rejects proprietary lock-in.