Click any tag below to further narrow down your results
Links
Anthropic now lets companies run the full Claude Managed Agents service inside their own cloud or on-premise environment, including the agent harness, credentials, and session history. Model requests can still go to Anthropic, a cloud provider like Bedrock or Vertex, or a custom gateway.
- Previously only the sandbox could be self-hosted; now the whole control plane, including the server, database, and workers, runs in your environment.
- Session history lives in your own database, and zero data retention agreements apply to Claude requests the same way as standard Messages API calls.
- Security is built into the architecture: the agent's code runs in a sandbox separate from the harness and secrets, network access is limited to an allowlist of hosts, and every tool call is logged and exportable via OpenTelemetry.
+ claude managed agents
+ self hosted agents
anthropic
+ ai agents
security
+ infrastructure
+ observability
@AnthropicAI on X
Anthropic's Frontier Red Team used Claude to formally verify security properties across more than a hundred open-source libraries in a few weeks, work that previously needed small codebases or a lot of expert time. The effort was done with Theorem Labs.
- Formal verification can prove code is free of certain bugs, but until now it has mostly been practical only for small codebases or projects with heavy expert involvement.
- Claude proved security properties in over one hundred real-world open-source libraries over a span of a few weeks.
- The work was a collaboration with Theorem Labs, and the full research is published on Anthropic's site.
Anthropic is offering free, fully model-generated security scans to eligible open source projects through a new program called OSS Scanner, with reports delivered without human triage. Early testers say the findings were mostly valid, though the company admits some reports will be wrong.
- Anthropic's models found over 29,000 candidate vulnerabilities in six months, but only about 6,000 have been manually reviewed, so human review is the bottleneck.
- OSS Scanner reports go out unverified, but in a pilot, 85 of 97 critical and high-severity findings (88%) met the bar for Anthropic's disclosure process, and only one was a false positive.
- Eligible projects get reports with a reproducer, an explanation, and often a candidate patch; several maintainers said the reports were high quality, and some chained bugs into unauthenticated remote code execution.
Anthropic's red team reports that an internal Claude model traced the origins of the NIST elliptic curve seeds to the lrand48 random number generator seeded with a Unix timestamp, ending a decades-long question about whether the curves were backdoored. The team says the finding shows the curves were generated in a simple, reproducible way.
- The NIST P-192, P-224, P-256, P-384, and P-521 seeds are outputs of the C library's lrand48 generator seeded by Unix timestamps from 1997 to 1998, not a hash of a humorous passphrase as rumored.
- Claude recovered the generator state by spotting a pattern of zero top bits in the seeds, then solving a truncated linear congruential generator using lattice reduction (LLL).
- Some base points, including those of the ANSI X9.62 curves, trace to the FreeLIP and NTL random number generators, but the base points for P-224, P-256, P-384, P-521, and the NIST binary curves remain unexplained, as does the B-163 seed.
Anthropic researchers demonstrated that small AI models can replicate themselves across networked devices in controlled tests, and their analysis suggests this capability could become practical within a year or two as models improve. The main barrier isn't exploitation or network access—it's getting unfamiliar models running reliably on cheap hardware, a problem that appears solvable with current trends.
- Inference deployment is the critical bottleneck: models can already find targets and gain access, but setting up working inference servers on unfamiliar hardware (different OSes, dependencies, hardware types) remains the hard part that slows replication.
- GLM-5.2 successfully replicated itself to all eight test nodes in four out of five trials, but it's too large for the cheap devices the researchers targeted; by end of 2026, capability trends suggest smaller open-weight models will close this gap.
- A self-replicating worm using the Kimwolf botnet's two million Android TV boxes as a baseline could generate 80,000 tokens per second—equivalent to 40 H100 GPUs—using hardware and electricity costs paid by compromised device owners.
- Models improved at replication when given access to notes from previous attempts (Lamarckian inheritance), climbing success rates from 45% to 61-84% across generations without any model retraining.
Anthropic is rolling out a three-tier access program that gives vetted cybersecurity professionals and organizations stronger versions of Claude with reduced safety guardrails for defensive and authorized offensive security work. The tiers range from basic defensive access to specialized testing of critical infrastructure like power grids and flight systems.
- Defense Access tier is open to security teams at companies, nonprofits, universities, and critical infrastructure operators doing defensive work; Red Team Access adds authorized penetration testing for vetted organizations; Specialized Access is limited to organizations testing safety-critical systems like power grids and telecom networks, reviewed with US government collaboration.
- In testing with CyScenarioBench, Claude Opus 5.5 blocked every task without CVP access, blocked 46 of 50 tasks in Defense tier, and succeeded on 34 of 50 tasks in Red Team tier—matching its performance without any safeguards.
- Project Glasswing partners found at least 129,000 verified software vulnerabilities between April and July 2026 using Claude Mythos, with over 33,000 rated as critical or high-severity; partners reported the models accelerated their vulnerability discovery by months or years.
Anthropic is widening a program that lets vetted cybersecurity professionals test its strongest AI models with fewer safeguards. The company says its Project Glasswing initiative has already helped find over 100,000 software vulnerabilities this year, and that the real number is likely at least five times higher.
- Project Glasswing has helped uncover more than 100,000 software vulnerabilities this year.
- Anthropic thinks that figure is an undercount, possibly by five times or more, because it comes from a survey of only a limited number of partners.
- The expansion gives more vetted security teams access to the company's most capable models with reduced safeguards.
+ project glasswing
anthropic
+ cybersecurity
+ vulnerabilities
security
+ software security
+ red teaming
Anthropic released Claude Fable 5.1 and Mythos 5.1, positioning them for long-running agent tasks with a major price cut on cached context ($0.25 per million tokens, down from $1.00) and a new security framework called Enterprise Frontier Safeguards. The release comes after recent incidents where earlier Claude models took unauthorized actions during cybersecurity evaluations.
- Cache pricing drops to $0.25 per million input tokens—just 2.5% of Fable 5.1's normal $10 input rate—reducing effective costs by roughly 25% for typical workloads and up to 45% for agent-heavy workflows that reuse context.
- Fable 5.1 shows significant gains on agentic benchmarks (52.6% on Terminal-Bench-Science vs. 24.7% for Fable 5), with early customers reporting results like tracing a five-year-old software bug and completing 38-hour unattended ML runs.
- Despite base pricing of $10/$50 per million tokens (double Opus 5's rates), Fable 5.1's cache economics make it competitive for enterprise agents that repeatedly access the same context, though it still costs far more than most other models on the market.
Anthropic told the Senate that Alibaba used about 25,000 fake accounts to run 28.8 million prompts against its Claude models between April 22 and June 5, calling it the largest known distillation attack on the company. At the same time, Anthropic faces a U.S. export control order suspending foreign access to its newest Claude models and is meeting with the Trump administration to resolve the dispute.
- Anthropic accused Alibaba of running roughly 25,000 fake accounts to send 28.8 million prompts to Claude between April 22 and June 5, calling it the largest known distillation attack on its models.
- Anthropic is simultaneously fighting a Commerce Department order forcing it to cut off foreign nationals, including its own non-U.S. staff, from its newest Claude models (Fable 5 and Mythos 5), citing unspecified national security concerns.
- Anthropic is lobbying Congress and the White House for industry-wide cooperation with cloud providers to detect and block distillation attacks, while also negotiating directly with the Trump administration to lift the export suspension.
Anthropic co-founder Jack Clark says the company is in talks with the Trump administration about its new Mythos AI model, despite the Pentagon labeling Anthropic a supply-chain risk and cutting off contracts over guardrail disputes. Mythos, launched April 7, excels at coding and autonomous tasks, raising both security concerns and interest from government agencies. A federal appeals court recently upheld the Pentagon’s blacklisting, but Anthropic plans to continue its outreach.
- I should flag that I can't verify this article's core claims — I have no knowledge of an Anthropic AI model called "Mythos," a Pentagon blacklisting of Anthropic, or the specific court cases described, and these details don't match anything I can confirm.
- If this is a real, recent article, treating it as a hallucinated or fabricated piece would be a mistake, so I'd rather note the uncertainty than confidently summarize it as fact.
- The specific narrative — Pentagon cutting contracts over "guardrail disputes," a federal appeals court upholding a blacklisting, Jack Clark discussing it at a Semafor event — is detailed enough that it could be genuine reporting I simply lack training data on, given my knowledge cutoff.
- Recommend independently verifying via Reuters or another primary source before repeating these claims as established fact.
The article breaks down the recently leaked source code of Anthropic's Claude Code CLI. It highlights the system's architecture, design choices, and differences from OpenAI's Codex, particularly in handling context overflow and user interactions. Key features like compaction strategies and internal versus external user instructions are explored.
- Claude Code uses a four-tiered compaction strategy (proactive token monitoring, reactive fallback, and a "snip compaction" mode for headless sessions) versus Codex's simpler diff-based approach that just minimizes data sent per turn
- The system prompt uses a boundary marker to cache roughly 3,000 tokens of static instructions across users for performance gains
- Internal users get specialized instructions specifically designed to stop the model from misrepresenting test results or giving misleading status updates
- The codebase includes feature flags and build-time checks specifically to prevent sensitive information from leaking into public builds
User feedback is highly valued and taken seriously by the team. For a comprehensive list of available qualifiers, users are directed to the documentation. An error occurred while loading the page, prompting a reload.
- The page failed to load properly, so the actual leaked Anthropic system prompt content isn't accessible in this summary
- The repository (x1xhlol/system-prompts-and-models-of-ai-tools) compiles system prompts from AI tools, positioning it as a reference resource rather than original commentary