37,617 results · Most relevant
Andrey Rybka recommends giving each agent its own identity with short-lived, scoped credentials and no shared keys, and killing any long-lived keys in favor of tightly scoped, short-lived, audience-bound tokens.
The speaker's rule of two for agent security: an agent may combine any two of untrusted input, sensitive data, and external action, but if all three are needed, human sign-off is required. He frames this as the priority to start with today.
Andrey Rybka recommends putting circuit breakers on anything catastrophic, human approval for any high-value actions, scoped credentials limited to intended actions, and tenant isolation.
Andrey Rybka recommends locking down agents using the rule of two: every tool call goes through a deterministic broker, and humans sign off for any high-value actions.
Andrey Rybka recommends sandboxing the environment the agent runs in, setting default deny for outbound network traffic including DNS, and placing egress controls outside the sandbox. He stresses that default deny means everything, including DNS, and that a sandbox with no web access still needs DNS blocked because data can be smuggled through DNS lookups.
The speaker recommends killing any long-lived keys and replacing them with tight-scoped, short-lived, audience-bound tokens, one per agent, scoped to the tenant domain. This limits what a hijacked or social-engineered agent can access.
Andrey Rybka recommends that tools should not be able to wipe anything — use least-privileged tools, and when an agent has access to tools that can wipe or destroy data, remove those tools entirely rather than relying on prompts.
The speaker recommends implementing a deterministic invocation broker for tool calls: every tool call goes through the broker, with human approval for high-value actions and hard rate limits and budgets that halt agent actions rather than just alerting.
Andrey Rybka recommends implementing hard tenant isolation as strict as AWS accounts or GCP projects, with workload identity tightly scoped to the domain, data provenance tagging, and an audit trail. He also advises checking the tenant on every data query.
The speaker recommends pairing sandboxed runtimes with default-deny egress for all outbound traffic including DNS, and adding provenance tagging, output filtering, and circuit breakers. Egress controls should sit outside the sandbox, and a formal runtime is preferred.
Andrey Rybka recommends using deterministic controls outside the model, checking every call, and keeping a kill switch for agents. He also notes that governance prompts must be enforced by deterministic code.
Andrey Rybka recommends implementing agent-to-agent mutual authentication because low-privileged agents should not be able to talk to higher-privileged agents, and using scoped credentials, least-privileged tools, and an audit trail for agent-to-agent interactions.
The speaker recommends circuit breakers on anything catastrophic-capable, wired into credential and egress paths so that when an agent breach occurs, a tripped breaker prevents exfiltration. He also recommends maintaining an audit trail and keeping guardrails outside the model.
Andrey Rybka recommends using a deterministic invocation broker and human approval for high-value actions like token transfers, noting that even a $5 transfer threshold is probably too much to allow without human approval.
Andrey Rybka recommends using supply chain verification, delaying ingestion of new packages on some mergers, and scoped credentials. He elaborates: don't just do supply chain verification; also delay ingestion — don't ingest every package on day zero, maybe wait a week or two.
Andrey Rybka recommends using injection protection, isolation, data separated from instructions, provenance tagging for all data, egress controls, and output filtering to defend against prompt injection and data exfiltration.
Farza's standard advice for people asking how to start making videos about their company: take out your iPhone, click record, selfie mode, talk about what you're doing for 30 seconds, and post it to Instagram, TikTok, Reddit, YouTube, or wherever. Only about 1-2% of people actually take him up on it.
To differentiate in a crowded market, go nichier and nichier. If you like jigsaw puzzles and are also a micro-brew connoisseur, your point of differentiation is combining both — doing the jigsaw content while also talking about a Rogue IPA. Often the beer part takes off and that's where the money gets made, not the jigsaw — or both, or you become the only beer-jigsaw person.
hypothetical scenario for the outcome
Carmen Li recommends verifying provider claims before purchasing by checking UIDs and inventories and running performance benchmarks such as NVIDIA DGX, InferenceX, and MLperf. For forward capacity, she says providers' purchase orders and permits should be reviewed before committing to future delivery.
To audit your AEO readiness, give the agent your websites and a few AEO posts from X/LinkedIn, ask it whether the tips are useful, and have it prompt different LLMs cold (logged out) to produce an AEO research score. In the SaaStr case, cold-prompted agents surfaced the company as an answer 46 out of 50 times, but answers vary across agents — so expect discrepancies and test repeatedly.
partial
Choose sources → to get a library of your own and a daily brief.