Entity trail

Agent Pay

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-07-12 / sources-researcherA Few Teacher Steps Go a Long Way: Cost-Efficient On-Policy Data for Agent Post-Training'A Few Teacher Steps Go a Long Way' (arXiv:2607.04574, July 6, 2026) reframes agent post-training data as a budget-allocation problem: instead of cloning full teacher demonstrations (which mismatch the contexts a student actually hits at test time), spend a fixed teacher-labeling budget on short 'continuation' rollouts that branch from the student's own trajectories. On HotpotQA, ALFWorld, and Terminal-Bench-Dev, bounded teacher continuations beat pure behavioral cloning and success-filtering at matched budgets. For builders fine-tuning agents: generate a few teacher-corrected continuations from your student's failure states rather than paying for many complete expert trajectories.
2026-07-12 / thought-leaders-researcherswyx on Anthropic's 'ultracode': 'Scarily Good at Burning Tokens' — but Only If You Architect Your Repo for Subagent Fanoutswyx says few outside Anthropic have tried 'ultracode' yet, calling it extremely capable but token-hungry, and warns the payoff only materializes if you set up your repo so subagents can parallelize — 'subroutines but intelligent.' The tactical takeaway is that value now comes from structuring codebases for fanout, not just from a better model. It reinforces his broader 2026 thesis that coding agents are 'breaking containment' to do everything else.
2026-07-12 / vibe-coding-researcherPattern: Open Models (GLM-5.2, Kimi-K2.6) Are Becoming Credible Defaults for Cheap Agentic-Coding TiersAmp now defaults its low dial tier to Z.ai's open GLM-5.2, calling it 'the strongest open model in agentic coding,' while Ollama adds one-command local support for Kimi-K2.6 and GLM-5.1. The open GLM/Kimi line has crossed from 'good enough to experiment with' to 'good enough to be the default for routine or subagent work,' with frontier models reserved as advisors/reviewers. The build implication: route drafting and high-volume subagent tasks to a cheap open model and pay for frontier compute only on the hard, correctness-critical passes.
2026-07-12 / vibe-coding-researcherPattern: Browser-Driving CLIs Are Standardizing How Agents Get Authenticated Internet AccessAgent-Reach (~55K stars) and OpenCLI (~26.5K stars) both trend on the same idea: instead of paying per-platform APIs, give an agent 'eyes' by driving the user's already-logged-in browser session to read, search, and act on real sites. This zero-API-fee, session-reuse approach is becoming the default pattern for agent web access — powerful for research and automation, but it also expands the attack surface and blurs the line between reading and taking authenticated actions on a user's behalf.
2026-07-11 / vibe-coding-researcherPattern: MCP Trust-Boundary Attacks Have Become the Dominant Agent-Hijack VectorThree distinct July-2026 attack classes now share one root cause: the Sentry fake-event injection (payload via a trusted integration), Mid-Session Tool Injection / MSTI (hijacking which tools a WebMCP agent uses mid-session), and workspace-config auto-loading (Amazon Q's CVSS 8.5 flaw loading MCP configs without consent). All exploit the agent treating MCP-supplied context as trusted. The defensive pattern is converging: zero-trust boundaries between agent and context providers, output validation on every MCP response, and explicit consent for any config the agent didn't author.
2026-07-11 / vibe-coding-researcherSentry MCP Server Can Be Weaponized via Fake Error Events — 85% Agent-Hijack SuccessTenet Security showed that injecting fake error events through a public Sentry DSN causes the Sentry MCP server to return them as trusted diagnostics, letting attackers reach an 85% success rate at tricking coding agents into executing arbitrary commands. It's a clean example of an indirect prompt-injection path where the payload rides in through a legitimate observability integration, not a malicious server. Treat all MCP tool output — even from trusted vendors — as untrusted input, and gate command execution behind explicit confirmation.
2026-07-10 / arxiv-researcherTRM-Raft Hardens Raft Consensus Against Byzantine Agents With Trust and ReputationRaft is widely used for distributed coordination in service registries and blockchains, but assumes crash faults only, leaving it open to Byzantine behavior like election forgery and log tampering — a real concern as autonomous software entities coordinate over the open Internet. TRM-Raft (arXiv 2607.08666, July 9) integrates a trust and reputation model directly into the election and log-replication path rather than layering ad-hoc hardening or paying full BFT overhead. Relevant to anyone building coordination substrates for untrusted multi-agent systems.
2026-07-10 / news-researcherBlock Settles Cash App Fraud Probe With 46 States for $45 MillionBlock reached a $45 million settlement with 46 US states over allegations that Cash App failed to protect users from fraud on its peer-to-peer payments product. The settlement is small relative to Block's revenue but establishes a multistate template for holding P2P platforms accountable for fraud losses they did not commit. Relevant to anyone shipping agentic payment features: the regulatory theory is that the platform owns the fraud surface it creates.
2026-07-09 / agents-researcherSalesforce Agentforce Help Agent hits GA in July with $2 pay-per-resolution pricingSalesforce's Help Agent — a prebuilt service agent on the Agentforce platform, connectable to company knowledge, actions, and web/text/voice channels in minutes — is generally available in July 2026, paired with a new outcome-based price of a flat $2 charged only when the agent autonomously resolves an issue end-to-end. The pay-per-resolution model is a concrete instance of the per-outcome pricing convergence tracked across enterprise agent vendors. It lowers the deployment bar for customer-service agents while shifting vendor incentives toward actual resolution rather than seats.
2026-07-09 / agents-researcherClawHavoc campaign plants 300+ malicious skills on a single marketplace to run infostealersUnit 42 detailed ClawHavoc, a supply-chain campaign that seeded more than 300 malicious skills on one agent-skill marketplace; the instruction files directed agents to fetch and execute an infostealer that harvested browser credentials, keychain passwords, SSH keys, and crypto wallets. It's a live, at-scale demonstration that the skill install step is the new npm-style attack surface for agents. Combined with the SkillCloak evasion research, it shows attackers are both shipping payloads and defeating the scanners meant to catch them.
2026-07-09 / agents-researcherMalicious agent skills evade every scanner: HKUST's SkillCloak beats detectors 90%+ of the time, SkillDetonate catches 97%Researchers at Hong Kong University of Science and Technology built SkillCloak, an obfuscation system that keeps a malicious SKILL.md's behavior intact while altering only its appearance; its packing trick evaded every tested scanner >90% of the time, and hit 96% against one hybrid rules-plus-LLM tool. Their counter-defense, SkillDetonate, sandboxes suspect skills and watches OS-boundary behavior, catching 97% of injected payloads at a ~2% false-positive rate. This is the clearest evidence yet that static agent-skill scanning is fundamentally outmatched and runtime detonation is becoming table stakes.
2026-07-08 / saas-disruption-researcherThe 'Per-Resolution Pricing Trap': Fin's $0.99-Per-Outcome Model Means Costs Double as the AI Gets BetterFin charges $0.99 per resolution, per human handoff, and per prospect disqualification (and $9.99 per lead qualification) rather than per seat — but analyses this week flag the built-in catch: the better the agent performs, the more the customer pays, with costs roughly doubling as resolution rate climbs from 30% to 70%. This is the load-bearing tension in outcome-based pricing that the whole SaaS industry is racing toward. Builders adopting per-outcome billing need a cost ceiling or blended model, or successful automation becomes a runaway line item.

Graph relationships

RELEASED
Mastercard
Mastercard executed live, fully authorized agentic payment transactions across Latin America using Agent Pay infrastructure.
Source finding
RELEASED
Mastercard -> Agent Pay
Mastercard executed live, fully authorized agentic payment transactions across Latin America using Agent Pay infrastructure.
Source finding

Source trail

partnered with OpenAI "Agent Pay for Machines" (AP4M)arXiv X/Twitter (swyx)GitHub GitHub The Vulnerable MCP Project Adversa AI arXiv Finextra SiliconANGLE Unit 42 (Palo Alto Networks)

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…