Hi,

Sent from my iPhone

> On Jun 23, 2015, at 2:32 AM, Randy Presuhn <[email protected]> 
> wrote:
> 
> Hi -
> 
>> From: "Black, David" <[email protected]>
>> Sent: Jun 22, 2015 5:06 PM
>> To: Randy Presuhn <[email protected]>, Kathleen Moriarty 
>> <[email protected]>
>> Cc: "[email protected]" 
>> <[email protected]>, 
>> "[email protected]" <[email protected]>, 
>> "[email protected]" <[email protected]>, The IESG <[email protected]>, 
>> "[email protected]" <[email protected]>, 
>> "[email protected]" 
>> <[email protected]>
>> Subject: Re: [OPSAWG] Kathleen Moriarty's No Objection on 
>> draft-ietf-opsawg-vmm-mib-03: (with COMMENT)
>> 
>> Hi Randy,
>> 
>>> I still don't see how monitoring a not-necessarily-
>>> trusted VM is in any way different from monitoring any other
>>> not-necessarily-trusted resource.  
>> 
>> IMHO, it's not, but that does not appear to be the source of the concern.
>> Here's an attempt to explain what I think is going on based on the AgentX
>> example:
>> 
>>> If one uses AgentX as the implementation technology, the protocol
>>> provides the subagent no means of access to other information known
>>> to the master agent.  
>> 
>> As I understand the emails the concern wrt AgentX would be:
>> 
>> If AgentX is used between:
>>    a) an SNMP engine plus master agent in a VM; and
>>    b) a sub-agent in the hypervisor,
>> then a malicious VM (that isn't supposed to be using AgentX) could
>> use AgentX to try to get into the hypervisor at least to access the
>> data in this MIB, and possibly for other nefarious purposes.
> 
> No, Agentx provides no way for a subagent to access MIB data.
> The worst that a malicious subagent can do is to impersonate
> a legitimate subagent.
> 
>> The secdir reviewer's original comment was:
>> 
>>> I would also mention the specific problem of software running in a virtual
>>> machine and accessing the hypervisor's variables. This is an attack vector
>>> that is somewhat specific to this MIB. It cannot be mitigated by network
>>> firewalls.
>> 
>> That would be a consideration if AgentX was running over some sort of
>> hypervisor syscall interface that could not be interposed on by a firewall.
>> In that scenario, some sort of access controls on that syscall interface
>> would be a good idea.
> 
> AgentX and the related protocols I'm aware of provide no mechanism
> for a subagent to access the master agent's variables.  

This is not clear in the draft.  If you make this clear, you put the problem 
out-of-scope and you might be providing helpful guidance to someone 
implementing with less experience than you.

This is a non-blocking comment, so I'll leave it to the Responsible AD.

Thanks,
Kathleen 

> Providing
> such a mechanism could completely negate the value of VACM.  If
> someone's homebrew interface allows this sort of wanton behaviour,
> that's their problem.  We can't (and should try to) call out every
> conceivable ill-advised implementation strategy.
> 
>> OTOH, if one forced the AgentX access into the hypervisor to be via network
>> traffic, then a firewall (e.g., a software firewall on the same hypervisor),
>> could provide some mitigation.
> 
> No.  It would not provide any more protection than any other IPC
> mechanism might.  AgentX and related mechanisms assume a degree
> of trust between the master and subagent.  This is documented,
> for example, in the AgentX RFC.  A consequence of this assumption
> is that the subagent code instrumenting a non-trusted resource
> should be implemented in a way that it is protected from misbehaviour
> by the resource; as a practical matter, this means in a separate
> subagent process, distinct from both the master agent and the
> stuff being managed.
> 
> None of these concerns are the least specific to management of
> hypervisors / VMs, and are, as far as I know, widely understood
> among implementors of subagent technologies.
> 
> Randy

_______________________________________________
OPSAWG mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/opsawg

Reply via email to