Black Hat starts Sunday in Las Vegas. The timing isn't a coincidence — several of this week's stories read like advance briefings for the conversations that are about to happen on the show floor. Here's what mattered.
Anthropic cut 80% of Claude Code's system prompt. Stop writing so many rules.
In a post published July 24, Anthropic described removing more than 80% of Claude Code's system prompt for its newest models — Claude Opus 5 and Claude Fable 5 — without losing ground on coding evaluations. That's not a minor edit. It's a signal about how context should be built for agentic tools going forward.
The company frames this as context engineering, distinct from prompt engineering. Where a prompt is specific to one request, context is assembled from system prompts, skills, memory, and other sources and gets reused across many requests. Anthropic's argument: the heavy rules and examples that made sense for earlier models can actually work against more capable ones. Newer models don't need as much hand-holding, and over-specifying instructions can restrict exploration rather than improve decisions.
The practical implications for anyone building agents: keep CLAUDE.md files focused on genuine codebase gotchas, not general coding knowledge. Design tools with clean, expressive parameters rather than burying instructions in examples. Load specialized context progressively — only when needed — rather than front-loading everything. Anthropic rolled out a claude doctor command and a /doctor command inside Claude Code to help teams audit and rightsize their instruction files. If you've been stacking up rules in these files for the past year, this is worth running.
The broader point applies well beyond Claude Code. As Mitch Ashley, VP and Practice Lead for Software Lifecycle Engineering at The Futurum Group, put it: "An agent's real policy is whatever context loads at runtime, which makes the loading path the control point. Audit what your agents see before the next model upgrade changes their judgment."
Read more
Versa's Zero Trust MCP inverts a core assumption about how MCP servers work
Most MCP implementations assume the server can reach out to whatever it needs — a database, an API, an internal system — using credentials passed through from the user. Versa's Zero Trust MCP removes that capability entirely.
Under Versa's architecture, an MCP server running in the cloud that needs to check a policy doesn't call that API directly. It sends the function signature back inside the network boundary, and the authenticated user's own client makes the call using permissions they already have. The MCP server never touches the asset. "Nobody on the outside has access to your system," said Sridhar Iyer, senior director of ML/AI at Versa Networks. "You don't need to provision anything extra for the agent."
The identity question follows directly: at Versa, agents don't get a separate identity. The permission boundary is the human's. Audit logs flag whether an action was automated, but the scope of what the agent can reach is still bounded by what the person who triggered it is allowed to do.
Iyer offered a concrete test for Black Hat: any vendor claiming to use an LLM to stop attacks on another LLM is greenwashing, since anyone who can jailbreak the target model can jailbreak the guard. Look for deterministic controls backed by explicit policy instead.
Read more
An AI agent found MFA backup codes in a Downloads folder. 1Password is fixing that.
The scenario isn't hypothetical: an AI agent with file system access scans a developer's machine, finds a plain-text backup codes file sitting in Downloads, and exfiltrates it. 1Password's response is Extended Access Management for AI agents — a framework that treats agents as non-human identities with their own credential lifecycle, scope controls, and audit trail, separate from the human accounts they operate on behalf of. With non-human identity proliferating faster than governance can track, this is exactly the kind of purpose-built tooling the category needs.
Read more
Cursor puts a price tag on every AI coding decision
Cursor's new router now shows the cost of every model call in real time — token consumption, latency, and model selected — directly in the workflow. For teams that have been flying blind on what AI coding assistance actually costs per task, this is the visibility layer that should have shipped earlier. It also positions Cursor to make smarter automated routing decisions as model pricing continues to shift.
Read more
PQC migration: what's actually happening in the field
A combined view from Accenture's Tom Patterson and Versa's Anusha Iyer on post-quantum cryptography migration timelines, blockers, and what organizations are actually doing versus what they say they're doing. If PQC is on your radar — and it should be, given the CISA SBOM guidance this week and the general direction of federal mandates — this is grounded reading from two practitioners with real client exposure.
Read more
Also this week
- CISA's 2026 SBOM guidance adds hash requirements and AI coverage — if you're building or shipping software in regulated environments, this one has deadlines attached. Read more
- GitHub and PyPI are betting on time to slow software supply chain attacks — specifically, a mandatory delay between package publication and installation that gives automated scanners a window to catch malicious packages before they spread. Read more
- Torq SOC Brain reasons from precedent rather than retrieving it — a meaningful architectural distinction in how AI handles security incident response. Read more
- Diffusion models are making a case that LLMs don't have to generate text one token at a time — worth a read if you're tracking the technical direction of the model layer. Read more
- CodePen 2.0 turns a design playground into a real deployment tool. Read more
- GitHub Copilot's cloud agent in Linear gets more team-level controls. Read more
The through-line this week: the governance conversation has caught up with the capability conversation. Context engineering, zero trust MCP, agent identity, SBOM requirements — these aren't future concerns. They're active design decisions for any team building with AI in 2026.
Black Hat coverage starts next week. See you then.
Developer Weekly Briefing is published every Friday on Coder Legion. Written by Tom Smith.